PLDC-Net: A Domain-Specific Base Model for Plant Leaf Disease Classification Domain Adaptation Tasks.

PLDC‐Net: A Domain‐Specific Base Model for Plant Leaf Disease Classification Domain Adaptation Tasks - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Plant Direct . 2026 Apr 17;10(4):e70167. doi: 10.1002/pld3.70167 Search in PMC Search in PubMed View in NLM Catalog Add to search PLDC‐Net: A Domain‐Specific Base Model for Plant Leaf Disease Classification Domain Adaptation Tasks David J Richter David J Richter 1 Department of Artificial Intelligence Convergence, Chonnam National University, Gwangju, South Korea Find articles by David J Richter 1 , Kyungbaek Kim Kyungbaek Kim 1 Department of Artificial Intelligence Convergence, Chonnam National University, Gwangju, South Korea Find articles by Kyungbaek Kim 1, ✉ Author information Article notes Copyright and License information 1 Department of Artificial Intelligence Convergence, Chonnam National University, Gwangju, South Korea * Correspondence: Kyungbaek Kim ( [email protected] ) ✉ Corresponding author. Revised 2026 Mar 27; Received 2025 Oct 29; Accepted 2026 Apr 2; Collection date 2026 Apr. © 2026 The Author(s). Plant Direct published by American Society of Plant Biologists and the Society for Experimental Biology and John Wiley & Sons Ltd. This is an open access article under the terms of the http://creativecommons.org/licenses/by-nc/4.0/ License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes. PMC Copyright notice PMCID: PMC13088867  PMID: 42005353 ABSTRACT Plant diseases are the cause of heavy losses of crop production and, therefore, a big contributor to food shortages. Identifying these diseases as early as possible is important to limit the negative effects that these diseases have on the yields, as slow response time will lead to the spread of diseases and further loss. Traditionally, trained staff will go into the fields, multiple times during the growth period, and inspect the plants in samples through field disease monitoring. These traditional processes are time‐consuming and costly, and can be error‐prone, if the staff is not properly educated or if the staff simply makes mistakes due to oversight, for example. To aid farmers with the process of correctly identifying diseases, artificial intelligence deep learning methods have been employed in recent years. However, to train such deep learning models, one needs to obtain sufficiently large and high‐quality datasets and a model architecture that is capable of extracting relevant features to accurately classify the plant leaves. Datasets are still a limitation in the field of plant leaf disease classification. As such, domain adaptation methods such as transfer learning are often employed to overcome this data shortage. However, in current research, these domain adaptation methods almost exclusively rely on ImageNet as the pretraining dataset, a dataset that is domain unrelated to plant leaf disease detection, and models are often left unmodified and un‐optimized as a result. In this work, we propose the pretraining of an improved attention‐based and SiLU‐activated DenseNet201 architecture called PLDC‐Net that is pretrained on a large‐scale plant leaf disease dataset constructed by the authors to create a domain‐specific base model for better domain adaptation to new plants and diseases, validating the improved results through transfer learning, fine‐tuning, one‐shot learning, and few‐shot learning. PLDC‐Net has managed up to just over 24% improvements in F1‐Score over the baseline in domain adaptation results. Keywords: base model, CNN, domain adaptation, few‐shot learning, one‐shot learning, plant leaf disease classification, transfer learning 1. Introduction Food security is still a problem that a lot of people suffer from worldwide. Currently, around 750 million people suffer from undernourishment, with a trend towards rising numbers (FAO et al. 2024 ). This equals roughly 9.1% of the population (as of 2023), a number that has risen from roughly 7.5% in 2019 (FAO et al. 2024 ). These numbers and statistics showcase that food security is an ongoing issue to this day, and that there is a heavy need to overcome it. The main source of calories and protein worldwide is crop products, with them contributing over 80% of calorie consumption and roughly 60% of the protein consumption (FAO 2023a ). As a result of that, between 2000 and 2021 crop production has seen an influx of about 54% (FAO 2023a ), with additional increases in demand of 50% being projected for the timeframe of 2012 to 2050 (FAO 2017 ). Based on these predictions and statistics, the importance of crop production and its yields on the global food supply becomes apparent, and the importance of assuring the continued production of crop products does too. Plant diseases and pests do, however, impact crop yields and therefore the production of crop‐based products (Iqbal et al. 2018 ). Every year, roughly 40% of the annual production is lost to such pests and diseases (FAO 2017 ). These high losses both impact the farmers and countries financially, and also the food supply as a result of losing this much of the potential possible yields. Early detection of diseases is very important to reduce the spread of diseases and reduce the loss of crops (FAO 2023b ). Traditionally, manual field scouting is carried out to detect diseases in the crop fields and contain their spread (Clay et al. 2012 ). This manual scouting and monitoring is carried out by trained personnel that goes out into the field, multiple times during the growth period (Koyshibayev and Muminjanov 2016 ), and samples plants in the field for whether or not they carry any diseases or not. All this needs to be done as fast as possible, since early detection is essential to minimizing losses (Europe fRO 2012 ). Recent advances in AI and deep learning (DL) methods have brought forward implementations of such DL models to the field of plant leaf disease classification. Images of plant leaves are fed to such models, often convolutional neural networks (CNN), and a diagnosis of the plant is returned by the CNN based on said input image (Khan et al. 2018 ). The training of DL models does, however, require large amounts of data (often labeled), which is not always readily available. In the case of plant leaf disease classification, a dataset would need to be collected by taking images of an individual plant's leaves, and then these leaves would have to be annotated with the disease they carry by an expert—a process that is time‐consuming, costly, and also requires the availability of enough plants of all classes (healthy leaves and leaves of all diseases one wants to be able to diagnose). Due to that, data are still an area in the field that requires more attention (Xu et al. 2023 ; Xu et al. 2024 ; Li et al. 2021a ). On top of that is the fact that datasets can be divided into three subcategories, those being lab, field, and hybrid datasets. Lab datasets are datasets that are collected in lab‐like conditions, one leaf at a time, removed from the plant on a simple background. Field datasets are the opposite, where images are taken in the field, where leaves are still attached to the plant and backgrounds are noisy, containing other plants, soil, etc. Hybrid datasets are the combination of the two, where the dataset contains images of both modalities. Lab datasets are most often used (Richter et al. 2025 ), but they often yield extremely high results in lab scenarios (Richter and Kim 2025a ) while not generalizing well to real‐life field conditions (Singh et al. 2020 ; Guth et al. 2023 ; Ahmad et al. 2023 ; Ferentinos 2018 ), where these models would most likely need to be used in application. Field and hybrid datasets generate more robust and generalizable models that can withstand different conditions and still perform (Guth et al. 2023 ; Ahmad et al. 2023 ; Ferentinos 2018 ). Field and hybrid datasets are, however, much more sparse, often not made public, and often smaller and less diverse than some of the big lab datasets (Richter et al. 2025 ), hence the current limitation of the dataset. To overcome such limitations in data, domain adaptation methods such as transfer learning (TL), fine‐tuning (FT), one‐shot learning (OSL), and few‐shot learning (FSL) are used. These methods use a pretrained base model, a model that was trained on a dataset that is not the target dataset, to learn to extract relevant features, before being domain adapted to the final target domain. This can overcome small data issues such as overfitting and biases. However, this pretraining is almost always carried out on the ImageNet dataset (Richter et al. 2025 ). The ImageNet dataset is huge (1000 classes and around 1.2 million images), but very general. It is not related to any specific domain and can therefore not learn features that are domain‐specific to any given field. This is also true for plant leaf diseases. ImageNet (Deng et al. 2009 ) does not contain any such images, and therefore the image extractor trained using ImageNet will not contain any specific features that would be relevant for this task. While a general feature extractor trained on ImageNet can still produce valuable features that can help to combat overfitting and other issues, domain‐specific pretrained base models offer better feature extractors that can produce better final results postdomain adaptation, since the domain gap between pretrained data and target domain data is much smaller, and so this much smaller gap can be overcome much more easily (Chen et al. 2023 ; Quan et al. 2023 ). To further improve results, different foundation model architectures can be selected, all with differing success in the field of plant leaf disease classification (Richter and Kim 2025a ), and these models can be modified to extract better features using additional advanced methods such as improved activation functions and attention mechanisms that can improve the model's capabilities. Combining such an improved model and training it on a large‐scale hybrid dataset in order to create a plant leaf disease classification domain‐specific base model for domain adaptation tasks will allow future researchers and farmers to train models with less data and to do so less computationally heavy while obtaining superior results. In this work, we present our Plant Leaf Disease Classification Network (PLDC‐Net), a channel attention (CA) and SiLU activation enhanced improved DenseNet201 (Huang et al. 2017 ) architecture that was pretrained on PLDC‐80 (Richter and Kim 2025b ; Richter 2025a ), a dataset that was created by the authors of this work by combining, merging, selecting, augmenting, and balancing already existing open datasets into one large and diverse multiplant, multidisease, and multicondition benchmarking and pretraining dataset. This domain‐specific pretrained base model will then be compared to an ImageNet pretrained DenseNet201 model in domain adaptation tasks (TL, FT, OSL, and FSL) on separate, before unseen, cross‐domain datasets (with no overlap in images and plant–disease combinations) to showcase the improved performance of PLDC‐Net over the baseline. As such, the contributions of this paper lie as follows: An improved SiLU and CA enhanced DenseNet201‐based CNN architecture for plant leaf disease classification tasks Pretraining of said model on a large‐scale plant leaf disease benchmarking dataset combined, merged, filtered, augmented, and balanced by the authors Analysis of the pretraining results compared to the baseline as well as ablation studies of the added mechanisms Domain adaptation of the PLDC‐80 pretrained PLDC‐Net on a domain‐related dataset with no overlap (in terms of plant–disease combination classes and images) using: • TL • FT • OSL • FSL A methodology combining the above mentioned approaches to train better performing models with small training data through domain‐specific pretraining and an improved CNN architecture The rest of this paper is structured as follows: In Section 2 recent work in DL for plant leaf disease classification will be reviewed. Section 3 will provide background information about the topics and methods discussed in this paper, while Section 4 will introduce the methodologies used in this work in detail. Section 5 will present the final PLDC‐Net model architecture and present and explain the results obtained during the experiments. Then Section 6 will discuss the findings of this work, with Section 7 focusing on limitations and Section 8 presenting possible future work, before Section 9 will conclude this paper. 2. Related Works The field of DL for plant leaf disease classification has seen attention and work in recent years. Winiarti and Pujiyanta ( 2024 ) train different CNN architectures, including VGG16, MobileNetV2, and a basic CNN model on a dataset that contains in‐field chili leaf images openly available online and images collected by the authors. The dataset contains 250 images across five classes (healthy, yellowish, white fly, leaf spot, and curl leaf). The models, after training using TL, reached scores of 88% for VGG16, 94% for MobileNetV2, and 91% for the custom basic CNN. Mazumder et al. ( 2024 ) train a wide range of ImageNet pretrained CNN models (MobileNetV2, DenseNet121, DenseNet169, DenseNet201, NASNetLarge, ResNet152V2, InceptionV3, InceptionResNetV2, EfficientNetV2S, and EfficientNetV2L) as well as a modified DenseNet201Plus on 2 open in field datasets, one of which is banana leaf images, while the other one contains black gram images. The datasets were augmented using scaling, rotation, flipping, noise, and cropping to increase the number of images. The banana dataset has three classes, while the black gram dataset has five. After 50 epochs with early stopping, the enhanced model reached 99.5% accuracy on the black gram dataset and 90.12% on the banana dataset. Shah et al. ( 2024 ) collected their own dataset of grape leaves in field in India. The resulting dataset, which is not shared with the public, contains a total of 1600 images post‐augmentation, which span across four classes. The models (VGG16, VGG19, Xception, ResNet50V2, InceptionV3, and InceptionResNetV2) were pretrained on ImageNet and then trained via TL on the mentioned dataset. After only 10 epochs of training, the ResNet50V2 model returned the best results with 98% accuracy, with VGG19 and Xception reaching the second‐best score with 96%. Haikal et al. ( 2024 ) present results of training three CNN models (ResNet34, MobileNetV3, and GoogleNet) on the paddyDoc dataset, a hybrid dataset of rice leaf images. The dataset was augmented, and the effect of the different augmentation techniques (traditional methods, mix‐based methods, and no augmentation) were assessed. The dataset, with over 10,000 images of four classes, was used for TL on the before mentioned models, for 100 epochs. Results showcased that, when using only field data, the mix‐based methods clearly improved the performance of the models, but when using lab data, the mix‐based results were not as good. Overall, the different augmentation techniques did however produce better results when compared to not using any augmentation. Hang et al. ( 2024 ) inspect the performance of a set of pretrained CNN models (MobileNetV2, MobileNet, NASNetMobile, and Xception) and fuzzy ensembles of the models. The dataset was collected by the authors in field in China and not shared publicly. The soybean dataset was then expanded by generating synthetic images using CycleGAN, which resulted in 8456 images belonging to seven classes. Each of the models were pretrained on ImageNet and then trained for 1000 epochs on the soybean dataset through TL. The results showcased that the fuzzy ensemble did manage to outperform the individual models. Salam et al. ( 2024 ) propose an approach for the classification of mulberry leaf diseases. To do this, they first collected a dataset in field in Bangladesh, a dataset of 1091 images prior to augmentation and 6000 post‐augmentation. The models, which were VGG19, ResNet50, and MobileNetV3Small, were pretrained on ImageNet before being applied to the collected mulberry dataset. On that dataset, the models were trained for 60 epochs, using k‐fold TL training, which resulted in MobileNetV3Small achieving the highest metrics, but with more variance than VGG19 and ResNet50. Hammou and Boubaker ( 2022 ) utilize DenseNet169 and InceptionV3 on the PlantVillage dataset. In this work, not all classes and images of PlantVillage are considered, but rather only the classes containing tomato leaves. This results in a dataset of 10 classes and 18,162 images. Both models used were pretrained on ImageNet before being exposed to the tomato leaf images during training. Training was carried out for 100 epochs, after which the models managed 100% accuracy. Li et al. ( 2021b ) showcase results for RegNet, SuffleNet, MobileNetV3, and EfficientNet‐B0 trained on an apple leaf dataset. The dataset was collected by the authors in field in China and includes five classes which total 3131 images. The models were all pretrained on ImageNet before being trained for 50 epochs on the apple leaf dataset, after which RegNet using the Ranger optimizer managed the best performance with 99.9% accuracy. Reda et al. ( 2022 ) present a mobile app for plant leaf disease classification. As part of the development of said app, the authors trained MobileNet, MobileNetV2, EfficientNetB0, and NasNetMobile on the PlantVillage dataset with 39 classes and 61,486 images. The models were pretrained on ImageNet before being trained with TL. Training was carried out for a total of 30 epochs, after which the EfficientNet model managed the best scores with 99% accuracy. Eunice et al. ( 2022 ) implement InceptionV4, VGG16, ResNet50, and DenseNet121 to be trained on PlantVillage. In this case, PlantVillage with 38 classes and 54,305 images was used. The models were, here too, pretrained on ImageNet, and TL was utilized to train the models on the plant leaf disease data. The models were trained for 30 epochs, after which scores of up to 99.87% (DenseNet121) were achieved. Sultan et al. ( 2025 ) utilize a slightly modified Xception model for their experiments in plant leaf disease classification. Xception was modified by adding a few more convolutional layers to the end. The Xception backbone was pretrained on ImageNet before the modified model was trained on plant data. The dataset is hybrid and includes fruits, vegetables, and flowers that are infected with fungal and bacterial diseases as well as with pests. Healthy leaves are also present. The modified Xception model managed to perform well in the experiments, outperforming comparative models. As becomes apparent, most researchers use ImageNet pretrained and unmodified models in their experiments. Only a few models were modified, and pretraining on data other than ImageNet is even less common. However, both of these methods can help to generate better performing models in the end, especially in cases where data are not only lab data but also data that are taken in more real‐world‐like conditions. 3. Background To automatically classify the plant leaf images into their respective diseases, we utilize DL CNN models in this work. A base model CNN, called PLDC‐Net, is introduced and trained on the PLDC‐80 dataset before it is used as a pretrained base for cross‐domain adaptation using TL, FT, OSL, and FSL. In this section, these mechanisms are explained. 3.1. CNN CNN are deep neural networks that are specialized for imaging tasks. CNN use kernels, which are often square in shape and made up of learnable parameters, that parse over the image, extracting features in the process. These learnable parameters get optimized during training to extract good, relevant, valuable, and important features that help the models classify images into their respective classes. This is an advancement over traditional approaches, where these features had to be coded by hand, a process that is time‐consuming, not easy to do, and also requires expert knowledge. These convolutional layers, which contain the kernels, are often stacked on top of each other, in combination with pooling layers (that reduce the image in size), to create CNN models. Earlier layers only see a small part of the image and often extract basic features, while later layers have a bigger receptive field and learn more complex features. This is why models that are too shallow often fail to generate good results, and models that are too deep tend to overfit to the training data. CNN see heavy usage in a wide array of tasks, and as such, they have seen advances through different new technologies that are added to the basic convolutional layer and pooling layer structure. 3.1.1. DenseNet The DeseNet architecture (Huang et al. 2017 ), for example, employs so‐called dense connections inside their dense blocks. These connections concatenate the input to the block to the output of it, which allows it to reuse earlier features, lowers the impact of the vanishing gradient problem, and speeds up the training of earlier layers. To allow the concatenation, all convolutions within a block are carried out with padding enabled, so the images are only downsized across blocks through pooling but do not get resized inside the blocks to maintain the concatenations. A single dense block contains a batch normalization (BN) layer, followed by a ReLU activation, a pointwise convolutional bottleneck reduction layer, another BN layer, another ReLU activation layer, before it computes the actual 2D convolutions with a kernel size of 3 × 3 (see Figure 1 ). DenseNet then stacks four of these dense block stacks on top of each other, with transition blocks in between them. Transition blocks contain the average pooling layer and downsize the image; the full architecture can be seen in Figure 2 . This architecture allows the training of strong models and of models that are well applicable to the field of plant leaf disease classification (Richter and Kim 2025a ). FIGURE 1. Open in a new tab The DenseNet block as it is used in DenseNet architectures. FIGURE 2. Open in a new tab The DenseNet201 model architecture. 3.1.2. Improved Mechanisms Certain techniques that can improve the performance of CNN models have since been developed. These include attention mechanisms that help the model focus on more important information while giving less focus to less important information and improved activation functions that help the converge more stable and perform better during training. 3.1.2.1. CA CA is a mechanism that is used to enhance CNN model performance. CA does that by helping the model distinguish between channels that carry less important features and channels that contain features that are very valuable to the task. It does that by using a CA block (see Figure 3 ), which weights channels based on their importance through trainable parameters that get optimized to select the best channels during training. This is accomplished by using global average pooling and global max pooling in parallel, which are then followed up by a separate fully connected encoder decoder structure, after which the parallel outputs get added and then multiplied to the input to the CA block, thus ranking the channels based on the learned added CA map. FIGURE 3. Open in a new tab The channel attention block. 3.1.2.2. SiLU Activation SiLU, or Swish‐1, activation (see Equation 1 ) is a self‐gated, non‐monochromatic, and nonzero negative activation function that is used in models such as the EfficientNet (Tan and Le 2019 ) family of models. When compared to the ReLU activation function, which is by far the most commonly used activation function in CNN, we can see quite a few differences (see Figure 4 ). While ReLU is 0 for all negative input values, SiLU is not, which helps minimize the effects of vanishing gradients. It is also non‐monochromatic, while ReLU is monochromatic, which allows for smoother training. Lastly, it is self‐gated, which allows for adaptive scaling. All these changes allow the SiLU function to improve performance during training. SiLU x = x · σ x = x 1 + e − x (1) FIGURE 4. Open in a new tab SiLU Swish‐1 activation compared to ReLU. 3.2. Base Models Base models are models with pretrained weights that are used in domain adaptation tasks. These models, through their pretrained weights, provide valuable information even prior to the start of the domain‐specific training that can help models adapt better, train faster, minimize overfitting, and limit biases. These models can be of many different architectures, and they can be pretrained on various different data sources, although ImageNet is most commonly used in imaging tasks. 3.3. Domain Adaptation Training CNN models requires large amounts of high‐quality labeled data, which is not always available. In cases in which data is sparse, researchers often utilize pretrained base models. These models were trained on a large‐scale dataset prior to training on the target domain dataset that the researchers plan to use. Through this pretraining on the pretraining dataset, the model can learn many features on the large‐scale pretraining dataset, and transfer and apply that knowledge to the new images in the target dataset. This allows the model to train faster, more robust, and potentially better models, with less risk of overfitting and bias. However, nevertheless, the model still needs to bridge the domain gap between the pretraining data domain and the target data domain. In most instances, pretraining is conducted on the ImageNet dataset, a dataset with over 1.2 million images across 1000 classes. This huge and diverse dataset allows the model to learn many general and robust features to be used during the retraining, but ImageNet (Deng et al. 2009 ) contains many diverse and general classes including goldfish, geese, cockroaches, baseballs, bookcases, candles, CD players, and many more. Most of these classes have little to no relation to the field of plant leaf disease classification. While many general features are still usable, obviously, many of the more nuanced and fine‐detailed features that are directly related to plant leaf diseases are not learnable through ImageNet training. As such, the gap between the pretrained data and the target data is rather large. Training on a dataset that itself already contains classes of plant leaf disease images can generate a base model with pretrained weights that are much more related and much closer to the target domain, shrinking the gap. This domain‐related pretraining allows models to further and better benefit from the pretraining and allows better results in terms of training time, computational requirements, and performance (Chen et al. 2023 ; Quan et al. 2023 ). To overcome the cross‐domain gap, no matter how big or small, different techniques are available. Some common methods include TL, FT, OSL, and FSL. 3.3.1. TL TL utilizes pretrained CNN by taking the weights of the feature extractor but applying them to the task at hand by replacing the classifier top with one that matches the class count required for the new task. Then during training, the classifier top can be trained to take the pretrained frozen and untrainable feature extractor's provided features and utilize them to predict the domain‐specific classes. This classifier top (which is trainable) can then bridge the gap off the pretrained feature extractor to the domain‐specific task. This allows the model to train fewer parameters and with less data than would otherwise be required as it can benefit from the pretrained features. 3.3.2. FT FT is a specialized TL category, where the feature extractor is not frozen but actually also trainable. As a result, it can be further adjusted to the target domain, which can produce better models, even with limited data, but it can also lead to more overfitting, more bias, and it is more computationally heavy. 3.3.3. OSL OSL is a method of creating a classification model on a pretrained base, with only a single image per class. This is done by not retraining any of the models' layers or parameters, but rather by cutting the model off at the end of the feature extractor, often after the global average pooling layer, and using it to generate embeddings. First, the one image per class is fed into the model, in inference prediction mode, after which the output of the global average pooling layer is saved and remembered as that class's embedding. Then, during testing or during regular usage, new images are fed into the system, their embeddings are extracted and then compared to the list of class embeddings via distance (often Euclidean distance, see Equation 2 ). Then the closest class embedding among all classes is the predicted class. d x , y = ∑ i = 1 n x i − y i 2 (2) This can generate new classifiers with very limited data—only one image per class and without much computational requirements—but also suffers from overfitting if target data is too diverse. As such, images for training should be representative of the conditions used in practice as much as possible. 3.3.4. FSL FSL uses the same concept that OSL implements, but simply with more images per class. Where OSL only has one image per class, FSL has a few, five in the case of this work. These five images are then fed into the CNN feature extractor just as they would in one shot, but here, after the embeddings are extracted, the embeddings are averaged across all images of the same class. As such, the embeddings are more representative and suffer less from overfitting than they would in OSL. 3.4. Plant Leaf Diseases Generally, one can group plant leaf diseases into three subcategories based on the pathogen. These three categories are fungal diseases, bacterial diseases, and viral diseases. Bacterial diseases, such as scabs, leaf spots, and fruit rots, for example, often enter the plant through natural openings in the plant and through wounds. They can then spread through seeds, via water, reproduce fast, and are hard to treat or even untreatable, so they should ideally be prevented (Bidlack and Jansky 2010 ; Penn State Extension 2012 ; Sarkar et al. 2023 ; Patel and Joshi 2017 ). Examples of bacterial diseases can be seen in Figure 5 . FIGURE 5. Open in a new tab Examples of plant leaf bacterial disease data obtained in lab. Tomato bacterial spot images obtained from PlantVillage Dataset (Mohanty et al. 2016 ). Fungal diseases (e.g., powdery mildew and blights) often spread through spores that can travel far through the air; cutting infected parts of the plants can reduce the spread of the fungal diseases (Penn State Extension 2012 ; Sarkar et al. 2023 ). Example images showing fungal diseases can be seen in Figure 6 . FIGURE 6. Open in a new tab Examples of pear plant leaf images taken in Field and of fungal disease images. Powdery mildew strawberry images obtained from the Strawberry Disease Detection Dataset (Afzaal et al. 2021 ). Viral diseases multiply in plants similarly to the way they do in humans, by hijacking the plants' cellular machinery. They spread through seeds and also insects. Commonly found diseases include mosaics, leaf rolls, and mop‐top (Penn State Extension 2012 ; Sarkar et al. 2023 ). 3.5. Datasets Datasets are integral to training good base models. Only a good dataset can train a strong base model, and a bad dataset (too small, not diverse enough, too low quality, etc.) can be detrimental to the robustness, performance, and generalizability of the model. A base model pretrained on good data, when used in cases where target domain data is sparse, can still generate good results by utilizing the features obtained from the pretraining dataset, so selecting the right dataset for pretraining is essential. As mentioned before, using domain‐related data can improve the capabilities of base models to better bridge the cross‐domain gap (Chen et al. 2023 ; Quan et al. 2023 ). But, to further help the model become more robust and to increase generalizability, using the right kind of dataset is important too. Generally, datasets in plant leaf disease detection can be grouped into three categories, those being field, lab, and hybrid datasets. 3.5.1. Lab Datasets Lab datasets refer to datasets that were captured under laboratory‐like conditions. This means that images of plant leaves are not taken in real‐world conditions that one would find in the field, but rather in conditions where leaves are captured, often one at a time with the leaf removed from the plant, on a perfect single‐color background. As such, the leaf in the image is void of any noise or crowded backgrounds, making the classification task much easier to handle, which leads to DL models reaching very high metrics when trained and tested on lab data (Richter and Kim 2025a ), but these circumstances lead to the model lacking generalizability in real‐world field conditions, where lab‐trained models tend to underperform (Guth et al. 2023 ; Ahmad et al. 2023 ; Ferentinos 2018 ), making them much less useful in real application use cases. Examples of images in lab conditions can be seen in Figure 5 . 3.5.2. Field Datasets Field datasets are essentially the opposite of lab datasets, as they are taken in field, with all the imperfections that come with that. Images in field datasets often contain noisy backgrounds or even noisy foregrounds, full of soil, other leaves, other plants, the photographers' hands, the sky, etc. This means that the model cannot, unlike in lab datasets, simply take the image and focus on the leaf right away but rather needs to learn what is important and what is not, to focus on the important features to predict. These types of images are of course more in line with what can be found in real‐world application use cases and therefore tend to generate models that can work better in tougher scenarios (Guth et al. 2023 ; Ahmad et al. 2023 ; Ferentinos 2018 ). Sample images showcasing field images can be seen in Figure 6 . 3.5.3. Hybrid Datasets Hybrid datasets are the combination of the two types of datasets mentioned above. Hybrid datasets contain images taken in lab‐like conditions as well as images taken in field‐like conditions. As such, the model is exposed to even more different image conditions, only leading to more robust models that can generalize even better than field‐only datasets (Guth et al. 2023 ; Ahmad et al. 2023 ; Ferentinos 2018 ). As such, when possible, it is recommended to use data that are as diverse as possible (in reasonable conditions) while still being high‐quality, large, and relevant. A hybrid dataset example is showcased in Figure 7 . FIGURE 7. Open in a new tab Examples of sugarcane plant leaf images taken in hybrid conditions and of viral disease images. Viral disease sugarcane images obtained from the Sugarcane Leaf Dataset (Thite et al. 2024 ). 4. Methodology To create PLDC‐Net, first, a baseline model architecture needs to be found, upon which the enhanced architecture can be built. Suitable enhancements need to be found before creating the final architecture of PLDC‐Net. That model then needs to be trained on the domain‐related pretraining dataset before being used in domain adaptation tasks to compare its performance to the baseline (DenseNet201 pretrained on ImageNet). 4.1. Dataset The dataset used in this work is the PLDC‐80 dataset, a new dataset that is a merged, curated, augmented, and balanced dataset based on nine original open datasets. These nine datasets can be seen in Table 1 . These datasets were merged, filtered, augmented (Bappi et al. 2025 ), and balanced to create a final dataset that is capable of training a base model with enough high‐quality images, enough diverse classes, enough diverse plants, and diseases in hybrid conditions. The final statistics for the PLDC‐80 dataset (Richter and Kim 2025b ) can be found in Table 2 , and plants and diseases are listed in Table 3 . TABLE 1. List of all the datasets that are included in PLDC‐80. References Dataset Plant Num. of class Total img. Type (Mwebaze et al. 2019 ) cassava Cassava 5 21,397 Field (Ahmad et al. 2021 ) cds Corn 3 1571 Field (Fenu and Malloci 2021 ) diaMOS Pear 4 3006 Field (Thapa et al. 2020 ) fgvc8 Apple 12 18,632 Field (Petchiammal et al. 2023 ) paddy Rice 10 10,407 Field (Liu et al. 2021 ) pdd271 Multi 10 7555 Field (Mohanty et al. 2016 ) plantVillage Multi 38 54,304 Lab (Afzaal et al. 2021 ) sms Strawberry 3 1583 Field (Thite et al. 2024 ) sugar Sugarcane 10 6405 Hybrid Open in a new tab TABLE 2. Key values of the PLDC‐80 dataset. Attribute Value Number of classes 80 Number of different plants 25 Number of images total 304,507 Train‐val‐test split 64–16–20 Number of images train 224,000 Number of images validation 56,000 Number of images test 24,507 Image size 224 × 224 × 3 Open in a new tab TABLE 3. Plants and diseases present in the dataset. Plant Corresponding disease Plant Corresponding disease Apple Black rot Rice Bacterial leaf blight Frog eye leaf spot Bacterial leaf streak Healthy Bacterial panicle blight Powdery mildew Blast Rust cedar apple rust Brown spot Scab Dead heart Bell pepper Bacterial spot Downy mildew Healthy Hispa Blueberry Healthy Normal Cassava Bacterial blight Tungro Brown streak disease Soybean Downy mildew Green mottle Healthy Healthy Squash Powdery mildew Mosaic disease Strawberry Angular leaf spot Cherry Healthy Healthy Powdery mildew Leaf scorch Corn Common rust Leaf spot Gray leaf spot Powdery mildew leaf Healthy Sugarcane Banded chlorosis Northern leaf blight Brown rust Northern leaf spot Brown spot Grape Black rot Grassy shoot Esca Healthy leaves Healthy Pokkah boeng Leaf blight Sett rot Leek Gray mold disease Smut Hail damage Viral disease Mung bean Brown spot Yellow leaf Orange Citrus greening disease Sweet potato Healthy leaf Peach Bacterial spot Magnesium deficiency Healthy Sooty mold Pear Slug Tomato Bacterial spot Spot Early blight Potato Early blight Healthy Late blight Late blight Radish Black spot disease Leaf mold Mosaic virus disease Septoria leaf spot Wrinkle virus disease Spider mites Raspberry Healthy Target spot Tomato mosaic virus Tomato yellow leaf curl virus Open in a new tab 4.2. Model To construct the PLDC‐Net model, a baseline architecture needs to first be chosen. For this, a set of five architectures that have proven to generate good results in plant leaf disease benchmarks (Richter and Kim 2025a ) was selected. These five architectures are ConvNeXt‐Tiny (Liu et al. 2022 ), ConvNeXt‐Small (Liu et al. 2022 ), EfficientNetV2‐B2 (Tan and Le 2021 ), EfficientNetV2‐S (Tan and Le 2021 ), and DenseNet201 (Huang et al. 2017 ). These model baselines were trained and tested on the PLDC‐80 dataset to benchmark and find the best model to build PLDC‐Net upon. Models were trained from scratch (no weights were loaded), and all parameters were being trained and updated during training. The model was chosen based on the performance on the test dataset, which only contains images unseen during training, providing a metric that ranks by generalization to new data. Each model was trained for 150 epochs, with the best performing instance at epoch n being reloaded after the 150 epochs have passed. All models were trained five times each, and the metrics were averaged across those runs to generate more representative results. The best performing baseline model will then be enhanced using suitable methods through an ablation study, in which the models get trained five times each, averaging the performance metrics, using the same dataset and hyperparameters as before (see Table 4 ). After evaluating the results and picking the best performing enhanced architecture (PLDC‐Net) and using its PLDC‐80‐pretrained weights, the domain adaptation evaluation will be carried out. Additionally, a comparative ablation study will be presented, in which the improvement mechanisms of PLDC‐Net will be applied to other basic CNN backbones—VGG16 (Simonyan and Zisserman 2014 ) with BN (Simon et al. 2016 ) and ResNet50 (He et al. 2016 )—to validate that the findings are relevant to the PLDC‐Net architecture specifically. All comparative experiments are run with identical hyperparameters and setup. TABLE 4. Hyperparameters and setup used during training. Setting Value Batch size 128 Epochs 150 Optimizer Adam Loss Categorical Crossentropy Callbacks Checkpoint (Val. Acc.) Machine CPU: Intel Xeon Gold 6330, GPU: NVIDIA A100 SXM4, RAM: 125 GB Open in a new tab 4.2.1. Hyperparameters and Setup Models were trained with the hyperparameters listed in Table 4 . Following established practice in the field, the Adam optimizer and Categorical Cross‐Entropy loss are adopted for training, which are widely validated for multiclass classification tasks and in the field of plant leaf disease classification (Richter et al. 2025 ; Richter and Kim 2025a ; Goodfellow et al. 2016 ). 4.3. Domain Adaptation Four different approaches of domain adaptation with PLDC‐Net as the base model will be tested. These are TL (with a frozen feature extractor), FT (with an unfrozen feature extractor), OSL, and FSL (with five images per class). They will be trained on in‐field images with noisy backgrounds, which are generally harder to handle but resemble real‐world use cases. Comparative domain adaptation experiments will be run using SOTA Vision Transformer (ViT) models (Dosovitskiy et al. 2020 ; Morales 2020 ), the B16 and L32 versions to be exact. The ViTs will also be pretrained from scratch on PLDC‐80 (on the same conditions) and will then, also under the same conditions as PLDC‐Net, undergo the same domain adaptation tasks. 4.3.1. TL and FT To train the baseline model (DenseNet201 pretrained on ImageNet) and the PLDC‐80 pretrained PLDC‐Net, a new dataset is needed that does not include any images seen during pretraining (no image overlap with PLDC‐80) and that does ideally not contain any plant–disease combination overlap (no class overlap) with the pretraining PLDC‐80, to guarantee true cross‐domain TL. As a result, two open datasets were combined and used to carry out these experiments (see Table 5 ), which resulted in the dataset shown in Table 6 . During the TL experiments, only the classifier head is trainable, while the CNN layers are frozen and untrainable. For FT, every layer is unfrozen and trainable. TABLE 5. The two datasets combined and used in transfer learning and fine‐tuning experiments. References Dataset Plant Num. of class Total img. Type (Lab 2020 ) iBean Beans 3 1296 Field (Mignoni et al. 2022 ) mignoniSoy Soybean 3 6410 Field Open in a new tab TABLE 6. Key values of the dataset used for transfer learning and fine‐tuning. Attribute Value Number of different plants 2 Number of classes 6 Number of images total 7706 Images per class Soybean healthy—896 Soybean caterpillar—3309 Soybean Diabrotica speciosa —2205 Bean healthy—428 Bean angular leaf spot—432 Bean bean rust—436 Image size 224 × 224 × 3 Open in a new tab This resulting dataset is heavily imbalanced, which resembles the types of dataset one would encounter when trying to generate a new dataset in real‐world use‐case scenarios, to truly test the applicability of the domain adaptation through TL and FT. To further simulate real conditions and to test the actual strength of the base models, train‐test splits will be set to be 10–90 (see Table 7 ). This will only give a small number of images during training to the base model, which makes training rely more on the base models, while resembling real data shortage in the field (Xu et al. 2023 , 2024 ; Li et al. 2021a ). Models will be evaluated on their F1‐Score due to the imbalance of the dataset. All experiments are run five times, and results are averaged to generate more representative results. PLDC‐Net is run in TL and FT once for each model obtained during the model pretraining resulting in five runs; the baseline is trained five times from the ImageNet weights baseline. Hyperparameters used during training can be found in Table 8 . TABLE 7. Train‐test splits for transfer learning and fine‐tuning experiments. Train‐test split (0.1–0.9) Train Test Soybean healthy 90 806 Soybean caterpillar 331 2978 Soybean Diabrotica speciosa 220 1984 Bean healthy 43 385 Bean angular leaf spot 43 389 Bean bean rust 44 392 Total 771 6934 Open in a new tab TABLE 8. Hyperparameters used for TL and FT. Setting Value Batch size Transfer learning: 128, fine‐tuning: 64 Epochs 30 Optimizer Adam Callbacks Checkpoint (test F1‐score) Loss Categorical Crossentropy Machine CPU: AMD Ryzen 7 5800X, GPU: NVIDIA GeForce RTX 3090, RAM: 64 GB Open in a new tab 4.3.2. OSL and FSL In the OSL and FSL experiments, only the iBean dataset (Lab 2020 ) will be used, since the iBean dataset does not contain augmented images and since the dataset is captured in more consistent conditions, allowing for training on such limited data. The resulting dataset values can be seen in Table 9 . For OSL, a single image will be used to generate the class embedding for each class for both models (PLDC‐Net and DenseNet201 with ImageNet weights). For FSL, five images will be given to generate embeddings, which are averaged per class for the five embeddings generated. All remaining images are used for testing. Class prediction is carried out through Euclidean distance to the class embedding, and test accuracy and F1‐Scores will be used as the main evaluation metric. The PLDC‐Net weights are not updated or trained at all during OSL and FSL, and only the embeddings will be predicted through forward runs through the pretrained PLDC‐Net. TABLE 9. Key values of the dataset used for One‐Shot. Attribute Value Reference Lab ( 2020 ) Number of different plants 1 Number of classes 3 Number of images total 1296 Images per class Healthy—428 Angular leaf spot—432 Bean Rust—436 Image Size 224 × 224 × 3 Open in a new tab 5. Experiments and Results In this section, the results of the experiments for finding the PLDC‐Net architecture for plant leaf disease detection are presented. Its results of training it on the PLDC‐80 dataset (Richter and Kim 2025b ), including comparison to other models that are well performing in the field (Richter and Kim 2025a ) and an ablation study of the mechanisms used, as well as domain adaptation experiments with PLDC‐Net as the base model are are shown and discussed. 5.1. Model First, the experiments for finding the most suitable improved model will be presented. 5.1.1. Baseline Model First, a baseline model to improve upon needs to be identified. This is done by comparing a set of models that are proven to perform well in the field (Richter and Kim 2025a ) (DenseNet201, EfficientNetV2‐B2, EfficientNetV2‐S, ConvNeXt‐Tiny, and ConvNeXt‐Small). They are trained on the PLDC‐80 dataset (Richter and Kim 2025b ), and their performance is compared to identify the best model. Test set accuracy is used to identify the best model. Results of the experiments can be found in Table 10 . All results are the average score of five iterations of the experiments to ensure more robust and more representative results. TABLE 10. Baseline model performances on PLDC‐80. Model Train acc. Val. acc. Test acc. ConvNeXt‐Tiny 99.360% 91.422% 84.290% ConvNeXt‐Small 99.412% 91.352% 84.220% EfficientNetV2‐B2 99.462% 94.464% 89.492% EfficientNetV2‐S 99.456% 94.696% 89.560% DenseNet201 99.382% 94.958% 90.536% Open in a new tab With the DenseNet201 architecture managing to outperform the other models by over 1% on the test data, the DenseNet201 model will be used as the baseline to improve upon in this work to build PLDC‐Net. 5.1.2. Model Improvement Ablation Study To improve the DenseNet201 baseline architecture, a set of improvements is implemented and then tested in an ablation study on the PLDC‐80 dataset to identify if they improve learning capabilities. The mechanisms used in this study to improve the DenseNet201 baseline are SiLU Swish‐1 activations and a CA block. The CA block is added after the last convolutional layers of the baseline DenseNet201, but before the global average pooling layer and the classifier top. The SiLU activation was added in place of all ReLU activations of the baseline DenseNet201 model, but not in the CA block. The CA block still uses ReLU. The model is trained five times for each configuration, and the average scores are presented and compared in Table 11 . A more extensive comparison of possible improvement mechanisms can be found in Richter ( 2025b ), the thesis to which this paper contributes. TABLE 11. Ablation study of the proposed model and its individual mechanisms on PLDC‐80. Model Train acc. Val. acc. Test acc. DenseNet201 99.382% 94.958% 90.536% DenseNet201 + SiLU 99.402% 94.842% 90.824% DenseNet201 + CA 99.446% 95.038% 90.896% PLDC‐Net (DenseNet201 + SiLU & CA) 99.400% 94.830% 91.054% Open in a new tab Results showcase that the DenseNet201 baseline is the worst, while both CA and SiLU improve upon it. The model with both added SiLU and CA performs the best during training, generating the best results on the unseen test data and being the only model that manages over 91% accuracy on said data on the five‐run average. As such, this architecture is chosen for its improved performance. This architecture with PLDC‐80‐pretrained weights will be referred to as PLDC‐Net moving forward and will be used as the plant leaf disease classification domain related base model in domain adaptation tasks. Figure 8 shows the final model's training performance over time. FIGURE 8. Open in a new tab Averaged training graph of the SiLU + CA DenseNet201 over the epochs. The shown graphs are the results on the validation set. Note that the model reached its peak performance after 124.6 epochs on average, to which the model would be reverted to at the end through early stopping. 5.1.2.1. Grad‐CAM Visualization Grad‐CAM (Selvaraju et al. 2017 ) is a visualization method for CNN that shows the activation map of convolution layers in the model. High activation areas showcase areas of the image that are important to the CNN at the current layer, whereas low activation areas bare little importance to the CNN. In this case, the activations are taken from the 11th block's last CNN layer, which is the second last convolutional block, as it is near the classifier head and therefore already has high impact on the actual classification, but still is high‐resolution enough to give valuable information on the spatial importance of features. The grad‐CAM visualization can be seen in Figure 9 . PLDC‐Net's generalization ability is showcased in Figure 10 , where images of plants and diseases that are not present at all during training are given to the model. One can see that even for plants and diseases foreign to the model, it manages to identify important features and areas due to the domain‐related pretraining. FIGURE 9. Open in a new tab Example Grad‐CAM images of the pretrained Channel Attention SiLU DenseNet201 model on before unseen test images of the PLDC‐80 dataset. FIGURE 10. Open in a new tab Grad‐CAM visualization of a PLDC‐Net on the OSL/FSL data. These are plants and diseases that were not present during training on an unchanged PLDC‐Net after only PLDC‐80 pretraining, showcasing PLDC‐Net's generalization ability. 5.1.2.2. Comparative Ablation Study To verify the validity of the proposed PLDC‐Net architecture as a whole, comparative ablation studies were carried out using two separate backbones. These are VGG16 (Simonyan and Zisserman 2014 ) and ResNet50 (He et al. 2016 ), since they both are CNN architectures that come without attention mechanisms and use ReLU activation by default. This allows for a fair and reliable ablation of the SiLU activation and the CA block. Note that VGG was enhanced slightly by adding BN to the first convolutional layer per block, as it would otherwise not be able to perform well in this scenario (Simon et al. 2016 ). All experiments in this ablation study were run twice and averaged. The results of these experiments can be seen in Table 12 . TABLE 12. A comparative ablation study of the proposed improvements to DenseNet201 (CA and SiLU, on other backbones). Model Train acc. Val. acc. Test acc. ResNet50 99.360% 93.645% 88.535% ResNet50 + CA 99.390% 93.715% 88.985% ResNet50 + SiLU 99.345% 93.095% 88.020% ResNet50 + CA + SiLU 99.350% 92.625% 87.335% VGG16 + BN 99.215% 90.460% 87.525% VGG16 + BN + CA 99.160% 90.875% 87.170% VGG16 + BN + SiLU 99.165% 90.900% 87.140% VGG16 + BN + CA + SiLU 99.185% 90.220% 85.405% Open in a new tab The results in Table 12 show that neither SiLU, CA, nor the combination of both managed to yield any improvements with the VGG16 + BN baseline, while only the CA block managed to slightly improve the ResNet50 results. One can also see that the best performing model for each backbone fails to outperform PLDC‐Net, which manages to generate significantly better results than both other backbones. Based on these results, one can see that these improvements are not an improvement to all backbones but rather work well in the proposed PLDC‐Net architecture specifically (see Table 11 ). 5.1.3. PLDC‐Net Architecture As a result of this, the resulting PLDC‐Net architecture looks as follows. The baseline model is DenseNet201. It is modified by using SiLU activation in place of all ReLU activations in the original DenseNet architecture. Also, after the feature extractor but before the global average pooling layer, the CA block is located, which still uses ReLU activation. The model can be seen in Figure 11 , and the code is openly available on GitHub (Richter 2025c ). FIGURE 11. Open in a new tab PLDC‐Net architecture. 5.2. Domain Adaptation PLDC‐Net will be tested in four different domain adaptation tasks to see if it functions as a better base model for this field than the baseline (unmodified DenseNet201 with ImageNet weights), due to it being improved in architecture and pretrained on domain‐relevant and related data. 5.2.1. TL Domain adaptation through TL will be done by freezing the feature extractor (all layers up until the global average pooling layer). This makes training less hardware‐intensive, since most layers and parameters do not need to be retrained. The PLDC‐Net (SiLU CA DenseNet201 with PLDC‐80 weights) will be compared to the baseline DenseNet201 model with ImageNet weights. Each of the five previously trained PLDC‐Net will be trained in this setting once, generating five sets of result data points, which will be averaged for better representation. The models were both trained on the dataset shown in Table 6 . Results can be seen in Table 13 . TABLE 13. Baseline and PLDC‐Net performance comparison in TL setting. All values are from the validation test set that the models were not trained on. Model Acc. F1‐score Prec. Recall Loss Baseline 72.410% 67.306% 74.124% 70.354% 0.8870 PLDC‐Net 79.454% 78.444% 83.722% 73.726% 0.5659 Open in a new tab As can be seen in the results, PLDC‐Net managed to outperform the baseline model in each metric, with accuracy scores higher by over 7% and F1‐Scores, which are arguably more important in an unbalanced dataset like this one, that are over 11% higher. This showcases that the domain‐related PLDC‐Net model allows for much stronger training of new domain‐specific models in plant leaf disease classification in limited data environments when compared to the baseline. The PLDC‐Net also managed to reach a higher F1‐Score after 10 epochs than the baseline did after all 30, showcasing it learns much faster (see Figure 12 ). The confusion matrix on the validation data can be seen in Figure 13 , and sample grad‐CAM images of the TL model can be seen in Figure 14 . FIGURE 12. Open in a new tab Performance comparison of average PLDC‐Net and baseline results in TL setting. FIGURE 13. Open in a new tab Confusion matrix of a TL domain adapted model on the validation data. FIGURE 14. Open in a new tab Grad‐CAM visualization of a PLDC‐Net after domain adaptation with TL. Images are from the validation set unseen during training. 5.2.2. FT FT, where both models are completely unfrozen and as a result are completely retrained, was carried out using the same dataset as during TL. This set of experiments has higher computational requirements than TL as a result of that but can also adjust more to the domain‐specific data due to this (but also is more at risk to overfit). The averaged results generated in these experiments, five for PLDC‐Net and five for the baseline, can be seen in Table 14 and Figures 15 and 16 . Figure 17 showcases an FT model's confusion matrix, and Figure 18 showcases grad‐CAM images after FT training. TABLE 14. Baseline and PLDC‐Net performance comparison in FT setting. All values are from the validation test set that the models were not trained on. Model Acc. F1‐score Prec. Recall Loss Baseline 64.090% 58.260% 64.284% 63.822% 4.17916 PLDC‐Net 82.884% 82.554% 83.912% 81.844% 0.61196 Open in a new tab FIGURE 15. Open in a new tab Performance comparison of average PLDC‐Net and baseline results in FT setting. FIGURE 16. Open in a new tab Zoomed‐out view of the losses in FT to visualize the domain gap, overfitting and volatility of the baseline during FT training. Note there here the y axis reaches 10,000, whereas before it went to 5. FIGURE 17. Open in a new tab Confusion matrix of an FT domain adapted model on the validation data. FIGURE 18. Open in a new tab Grad‐CAM visualization of a PLDC‐Net after domain adaptation with FL. Images are from the validation set unseen during training. The results show that PLDC‐Net manages to achieve significantly better results on all metrics in FT settings. The accuracy is over 18% higher, while the F1‐Score manages an improvement of over 24%. This shows very large improvements in the FT setting with low data, showcasing that the domain‐relevant PLDC‐Net is much better suited for the field than the general baseline. The PLDC‐Net model managed to learn much faster, achieving a better score after only two epochs than the baseline reached after 30. Also, when looking at Figure 15 and especially Figure 16 , it becomes evident that PLDC‐Net not only learns faster and better models but also learns much more robust and less volatile. 5.2.3. OSL In OSL, no back‐propagation DL learning of the model is performed, but rather the one image per class is used to generate embeddings per class, which the images to predict are then compared against. The dataset used here can be seen in Table 9 . PLDC‐Net embeddings are carried out once per pretrained model, and the results are then averaged across those runs for more representative results. Results can be seen in Table 15 . TABLE 15. Baseline and PLDC‐Net performance comparison in OSL setting. All values are from the validation test set that the models were not trained on. Model Acc. Recall Prec. F1‐score Baseline 53.95% 54.06% 56.12% 53.44% PLDC‐Net 59.27% 59.32% 67.99% 57.10% Open in a new tab Results showcase that PLDC‐Net is performing better in each metric, with an accuracy that is over 5% higher and a F1‐score that is almost 4% better. This proves that PLDC‐Net is able to consistently produce better performing models through domain adaptation using OSL in cases with extremely limited data. A confusion matrix of the OSL model's performance on the validation can be seen in Figure 19 . Grad‐CAM images of the PLDC‐Net on OSL data can be seen in Figure 10 . FIGURE 19. Open in a new tab Confusion matrix of an OSL domain adapted model on the validation data. 5.2.4. FSL FSL was carried out using the same dataset as OSL, but in this case, five images were used to create the class embeddings. The methodology is also largely the same, with the only difference being the larger training set. Results of these runs can be seen in Table 16 . TABLE 16. Baseline and PLDC‐Net performance comparison in FSL setting. All values are from the validation test set that the models were not trained on. Model Acc. Recall Prec. F1‐score Baseline 61.80% 61.83% 61.76% 61.78% PLDC‐Net 75.45% 75.52% 75.62% 75.42% Open in a new tab When looking at the results, we can again see that PLDC‐Net is better across all metrics. The accuracy is over 13% higher than the baseline, while PLDC‐Net also manages an F1‐score with over 13% better performance. This again showcases that PLDC‐Net is much better suited to be used as a base model in the field of plant leaf disease classification. The FSL confusion matrix can be seen in Figure 20 . Since OSL and FSL do not update the PLDC‐Net's weights, the grad‐CAM images for PLDC‐Net on the FSL data are also in Figure 10 . FIGURE 20. Open in a new tab Confusion matrix of a FSL domain adapted model on the validation data. 5.2.5. Comparative Domain Adaptation To further compare the results of the PLDC‐Net's domain adaptation performance, two ViT models (B16 and L32) were pretrained on PLDC‐80 in the same way that PLDC‐Net was and were then also, under the same conditions, used in domain adaptation tasks. Results of the ViT pretraining on PLDC‐80 can be seen in Table 17 , and the results of the ViT domain adaptation tasks are available in Table 18 . TABLE 17. Pretraining performacne of ViT models on PLDC‐80. Model Train acc. Val. acc. Test acc. ViT L32 57.698% 56.090% 56.500% ViT B16 61.012% 59.474% 61.412% Open in a new tab TABLE 18. Results of vision transformer models in domain adaptation tasks. DA stands for domain adaptation, and F1 is short for the F1‐score. Model DA Acc. Recall Prec. F1 ViT L32 TL 66.510% 52.658% 76.606% 54.506% ViT B16 TL 66.890% 52.842% 75.706% 56.128% ViT L32 FT 69.228% 61.608% 74.460% 62.126% ViT B16 FT 70.770% 65.286% 74.614% 63.660% ViT L32 OSL 46.596% 46.692% 46.426% 44.748% ViT B16 OSL 44.630% 44.684% 43.874% 42.764% ViT L32 FSL 50.748% 50.760% 50.658% 50.546% ViT B16 FSL 54.220% 54.276% 54.170% 53.812% Open in a new tab When looking at the results, it becomes obvious that ViT performs worse in all applications, which can most likely be attributed to the need for very large datasets that ViT requires to perform at their best (Dosovitskiy et al. 2020 ), whereas CNN can perform well on comparatively smaller data. This is true for pretraining, but even more so in the domain adaptation tasks, which were carried out in very low‐data scenarios, similar to how they can often be found in real‐world use cases. PLDC‐Net manages to constantly and heavily outperform both ViT architectures, proving PLDC‐Net's capability to generate good results in small data settings. 5.3. Model Parameter Comparison All models used in comparative studies, as well as the DenseNet201, its ablations, and PLDC‐Net are listed in Table 19 , alongside their parameter count, Giga Floating Point Operations (GFLOPS), their parameter MB allocation, and their peak GPU memory allocation. TABLE 19. Parameter comparison of the models used. Model GFLOPS Total params. Total MB GPU peak MB ConvNeXt‐Small 17.15 49,516,208 188.89 535.18 ConvNeXt‐Tiny 8.78 27,881,648 106.36 453.24 DenseNet201 8.58 18,475,664 70.48 146.09 DenseNet + CA 8.59 19,397,264 73.99 149.61 DenseNet201 + SiLU 8.63 18,475,664 70.48 146.09 EfficientNetV2‐B2 2.35 8,882,094 33.88 206.35 EfficientNetV2‐S 5.73 20,433,840 77.95 419.45 ResNet50 7.73 23,751,632 90.61 734.64 VGG16 + BN 30.71 14,761,616 56.31 2305.63 ViT B16 35.24 85,860,176 327.53 367.9 ViT L32 30.81 305,592,400 1165.74 1246.74 PLDC‐Net 8.64 19,397,264 73.99 149.61 Open in a new tab Table 19 showcases how heavy Vision Transformer models (ViT) are in comparison to CNN, which also explains their heavy reliance on very large datasets, hence their weak performance in the given experiments. The DenseNet models find a good balance between decent size and good performance. This table also shows that PLDC‐Net does not increase the size and computational complexity all too significantly when compared to the DenseNet201 baseline. 6. Discussion The results presented in this work showcase that PLDC‐Net (SiLU activated and CA DenseNet201 trained on the large‐scale PLDC‐80 plant leaf disease classification dataset) manages to outperform the baseline model in all four tested and observed domain adaptation tasks. In a field where data can be sparse and diverse, depending on the use case, traditional DL methods will be subject to overfitting and suffering from bias if trained from scratch. Methods like TL, FT, OSL, and FSL with a pretrained base model can help overcome some of these issues and therefore help in bridging the cross‐domain gap. However, due to the vast difference in image scope of ImageNet (which is almost exclusively used as the base model pretraining dataset in current research; Richter et al. 2025 ), the transferable knowledge can be limited to the point where the benefit too is limited. Using domain‐related data for pretraining can generate models that can bridge the domain gap more easily, faster and with less problems. Using such a domain‐related base model will allow the field of plant leaf disease classification to generate and train new models for specific purposes, be it new plants and crops, new diseases, new conditions, etc. As such the findings of this paper will allow researchers, and farmers, among others, to better train models for their specific purposes based on PLDC‐Net. 7. Limitations PLDC‐Net manages to achieve high accuracy scores during pretraining on the PLDC‐80 dataset and also on separate data after domain adaptation, aided by the domain‐related pretraining. However, there are some limitations that were observed that will be shared in this section. When inspecting the grad‐CAM visualization of some images, it can be observed that in some cases, the model is missing the truly important features and rather learns information that is unique to the class but not relevant to the disease, e.g., backgrounds or other image conditions that are only present in a single or a few classes (see Figure 21 ). While this occurrence is limited and not too common, overcoming it could further increase the model's performance. This issue could most likely be overcome if one could acquire or capture a large‐scale dataset in which all classes are taken in constant conditions (ideally in hybrid settings). Future experiments could also, if enough compute is available, test even more backbone architectures with more enhancement blocks, ideally with constant seeds and with multiple iterations per setup for better reproducibility and more representative data. FIGURE 21. Open in a new tab Grad‐CAM visualization of misclassifications of PLDC‐Net (backbone and all domain adaptations) and their grad‐CAM visualization that showcases the data inconsistency issue in some cases. 8. Future Work Future research in the field could use PLDC‐Net as a base model and adapt it to new domains (new plants, new diseases, etc.) and improve results over ImageNet pretrained models while testing the capabilities in new and diverse scenarios. Future work could also be conducted in the construction of the base model. One could further focus on the model architecture and try to find novel ways of further improving the model performance during pretraining. Additionally, one could experiment with different hyperparameters (e.g., batch size, optimizer, and loss), setups, or pretraining data to see if that can further improve results. Another possible direction would be to use other domain adaptation methods to bridge the domain gap from PLDC‐Net to the new target domains. Additionally, if possible, a truly new dataset of similar size to PLDC‐80, but all taken in constant conditions, would help the training of models in this field immensely. 9. Conclusion In this paper PLDC‐Net, a SiLU and CA enhanced DenseNet201 with plant leaf disease classification domain‐related pretrained weights, is presented. The model was pretrained on a large‐scale plant leaf disease classification dataset (PLDC‐80) and proven via ablation study. This domain‐related base model, with its improved architecture and domain‐related pretraining approach, allows faster, more robust, less volatile, and better performing training of domain‐specific domain adaptation of new models in the field. The results showcase that PLDC‐Net manages to outperform the baseline model in all four domain adaptation tasks significantly, with improvements of around 11% in TL, roughly 24% in FT, around 4% in OSL, and roughly 14% in FSL when compared to the baseline model. This proves that domain‐related base models can significantly help during the training of domain‐specific domain adaptation training in low‐data settings. Author Contributions D.J.R. generated the idea, designed the model, wrote the code, conceptualized and executed the experiments, collected and analyzed the data, wrote the main manuscript text, and prepared figures and reviewed. K.K. validated, reviewed, and advised. Funding This work was supported by Korea Institute of Planning and Evaluation for Technology in Food, Agriculture and Forestry (IPET) through the Agriculture and Food Convergence Technologies Program for Research Manpower development, funded by Ministry of Agriculture, Food and Rural Affairs (MAFRA) (project no. RS‐2024‐00397026, 34%). This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP)‐Innovative Human Resource Development for Local Intellectualization program grant funded by the Korea government (MSIT) (IITP‐2026‐RS‐2022‐00156287, 33%). This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) under the Artificial Intelligence Convergence Innovation Human Resources Development (IITP‐2026‐RS‐2023‐00256629, 33%) grant funded by the Korea government (MSIT). Conflicts of Interest The authors declare no conflicts of interest. Acknowledgments We appreciate the high‐performance GPU computing support of HPC‐AI Open Infrastructure via GIST SCENT. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. References Afzaal, U. , Bhattarai B., Pandeya Y. R., and Lee J.. 2021. “An Instance Segmentation Model for Strawberry Diseases Based on Mask R‐Cnn.” Sensors (Basel) 21, no. 19: 6565. 10.3390/s21196565. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Ahmad, A. , El Gamal A., and Saraswat D.. 2023. “Toward Generalization of Deep Learning‐Based Plant Disease Identification Under Controlled and Field Conditions.” IEEE Access 11: 9042–9057. 10.1109/ACCESS.2023.3240100. [ DOI ] [ Google Scholar ] Ahmad, A. , Saraswat D., Gamal A. E., and Johal G.. 2021. “CD&S Dataset: Handheld Imagery Dataset Acquired Under Field Conditions for Corn Disease Identification and Severity Estimation.” arXiv Preprint arXiv:2110.12084. 10.48550/arXiv.2110.12084. [ DOI ] Bappi, M. I. , Richter D. J., and Kim K.. 2025. “Assessing the Effectiveness of Augmentation Techniques in Enhancing Plant Leaf Disease Classification.” Smart Media Journal 14, no. 1: 17–25. 10.30693/SMJ.2025.14.1.17. [ DOI ] [ Google Scholar ] Bidlack, J. E. , and Jansky S.. 2010. Stern's Introductory Plant Biology. 12th ed. McGraw‐Hill New York. [ Google Scholar ] Chen, F. , Giuffrida M. V., and Tsaftaris S. A.. 2023. “Adapting Vision Foundation Models for Plant Phenotyping.” 604–613. Clay, D. C. , Carlson C. G., and Dalsted K.. 2012. iGrow Wheat: Best Management Practices for Wheat Production. South Dakota State University, College of Agriculture and Biological Sciences. SDSU Extension; Agronomy, Horticulture, and Plant Science Books. 1. [ Google Scholar ] Deng, J. , Dong W., Socher R., Li L. J., Li K., and Fei‐Fei L.. 2009. Imagenet: A Large‐Scale Hierarchical Image Database, 248–255. IEEE. [ Google Scholar ] Dosovitskiy, A. , Beyer L., Kolesnikov A. et al. 2020. “An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale.” arXiv Preprint arXiv:2010.11929. 10.48550/arXiv.2010.11929. [ DOI ] Eunice, J. , Popescu D. E., Chowdary M. K., and Hemanth J.. 2022. “Deep Learning‐Based Leaf Disease Detection in Crops Using Images for Agricultural Applications.” Agronomy 12, no. 10: 2395. 10.3390/agronomy12102395. [ DOI ] [ Google Scholar ] Europe fRO . 2012. Monitoring and Surveillance of Cereals Pests, Diseases and Weeds. 1st ed. Sub‐Regional Office of FAO for Central Asia: FAO; FAO Job Number: AQ295E. [ Google Scholar ] FAO . 2017. The Future of Food and Agriculture: Trends and Challenges. 1st ed. FAO. [ Google Scholar ] FAO . 2023a. World Food and Agriculture—Statistical Yearbook 2023. No. 2023 in FAO Statistical Yearbook—World Food and Agriculture. FAO. [ Google Scholar ] FAO . 2023b. The State of Food and Agriculture 2023—Revealing the True Cost of Food to Transform Agrifood Systems. No. 2023 in the State of Food and Agriculture (SOFA). FAO. [ Google Scholar ] FAO , IFAD , UNICEF , WFP , and WHO . 2024. The State of Food Security and Nutrition in the World 2024—Financing to End Hunger, Food Insecurity and Malnutrition in All Its Forms. No. 2024 in the State of Food Security and Nutrition in the World (SOFI). FAO, IFAD, UNICEF, WFP, WHO. [ Google Scholar ] Fenu, G. , and Malloci F. M.. 2021. “DiaMOS Plant: A Dataset for Diagnosis and Monitoring Plant Disease.” Agronomy 11, no. 11: 2107. 10.3390/agronomy11112107. [ DOI ] [ Google Scholar ] Ferentinos, K. P.

2018. “Deep Learning Models for Plant Disease Detection and Diagnosis.” Computers and Electronics in Agriculture 145: 311–318. 10.1016/j.compag.2018.01.009. [ DOI ] [ Google Scholar ] Goodfellow, I. , Bengio Y., and Courville A.. 2016. Deep Learning. MIT Press. http://www.deeplearningbook.org . [ Google Scholar ] Guth, F. A. , Ward S., and McDonnell K.. 2023. “From Lab to Field: An Empirical Study on the Generalization of Convolutional Neural Networks Towards Crop Disease Detection.” European Journal of Engineering and Technology Research 8, no. 2: 33–40. 10.24018/ejeng.2023.8.2.2773. [ DOI ] [ Google Scholar ] Haikal, A. L. A. , Yudistira N., and Ridok A.. 2024. “Comprehensive Mixed‐Based Data Augmentation for Detection of Rice Leaf Disease in the Wild.” Crop Protection 184: 106816. 10.1016/j.cropro.2024.106816. [ DOI ] [ Google Scholar ] Hammou, D. R. , and Boubaker M.. 2022. Tomato Plant Disease Detection and Classification Using Convolutional Neural Network Architectures Technologies, 33–44. Springer. [ Google Scholar ] Hang, Y. , Meng X., and Wu Q.. 2024. “Application of Improved Lightweight Network and Choquet Fuzzy Ensemble Technology for Soybean Disease Identification.” IEEE Access 12: 25146–25163. 10.1109/ACCESS.2024.3365829. [ DOI ] [ Google Scholar ] He, K. , Zhang X., Ren S., and Sun J.. 2016. “Deep Residual Learning for Image Recognition.” 770–778. Huang, G. , Liu Z., Van Der Maaten L., and Weinberger K. Q.. 2017. “Densely Connected Convolutional Networks.” 4700–4708. Iqbal, Z. , Khan M. A., Sharif M., Shah J. H., Rehman u M H., and Javed K.. 2018. “An Automated Detection and Classification of Citrus Plant Diseases Using Image Processing Techniques: A Review.” Computers and Electronics in Agriculture 153: 12–32. 10.1016/j.compag.2018.07.032. [ DOI ] [ Google Scholar ] Khan, M. A. , Akram T., Sharif M., et al. 2018. “CCDF: Automatic System for Segmentation and Recognition of Fruit Crops Diseases Based on Correlation Coefficient and Deep CNN Features.” Computers and Electronics in Agriculture 155: 220–236. 10.1016/j.compag.2018.10.013. [ DOI ] [ Google Scholar ] Koyshibayev, M. , and Muminjanov H.. 2016. Guidelines for Monitoring Diseases, Pests and Weeds in Cereal Crops. 1st ed. FAO. [ Google Scholar ] Lab, M. A.

2020. “Bean Disease Dataset.” Online. Li, L. , Zhang S., and Wang B.. 2021a. “Plant Disease Detection and Classification by Deep Learning a Review.” IEEE Access 9: 56683–56698. 10.1109/ACCESS.2021.3069646. [ DOI ] [ Google Scholar ] Li, L. , Zhang S., and Wang B.. 2021b. “Apple Leaf Disease Identification With a Small and Imbalanced Dataset Based on Lightweight Convolutional Networks.” Sensors (Basel) 22, no. 1: 173. 10.3390/s22010173. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Liu, X. , Min W., Mei S., Wang L., and Jiang S.. 2021. “Plant Disease Recognition: A Large‐Scale Benchmark Dataset and a Visual Region and Loss Reweighting Approach.” IEEE Transactions on Image Processing 30: 2003–2015. 10.1109/TIP.2021.3049334. [ DOI ] [ PubMed ] [ Google Scholar ] Liu, Z. , Mao H., Wu C. Y., Feichtenhofer C., Darrell T., and Xie S.. 2022. “A Convnet for the 2020s.” 11976–11986. Mazumder, M. K. A. , Kabir M. M., Rahman A., Abdullah‐Al‐Jubair M., and Mridha M.. 2024. “DenseNet201Plus: Cost‐Effective Transfer‐Learning Architecture for Rapid Leaf Disease Identification With Attention Mechanisms.” Heliyon 10, no. 15: e35625. 10.1016/j.heliyon.2024.e35625. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Mignoni, M. E. , Honorato A., Kunst R., Righi R., and Massuquetti A.. 2022. “Soybean Images Dataset for Caterpillar and Diabrotica speciosa Pest Detection and Classification.” Data in Brief 40: 107756. 10.1016/j.dib.2021.107756. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Mohanty, S. P. , Hughes D. P., and Salathé M.. 2016. “Using Deep Learning for Image‐Based Plant Disease Detection.” Frontiers in Plant Science 7: 1419. 10.3389/fpls.2016.01419. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Morales, F.

2020. “Vit‐Keras: Keras Implementation of Vision Transformer (ViT).” https://github.com/faustomorales/vit‐keras . Mwebaze, E. , Gebru T., Frome A., Nsumba S., and Tusubira J.. 2019. “ICassava 2019 Fine‐Grained Visual Categorization Challenge.” arXiv Preprint arXiv:1908.02900. 10.48550/arXiv.1908.02900. [ DOI ] Patel, A. , and Joshi B.. 2017. “A Survey on the Plant Leaf Disease Detection Techniques.” International Journal of Advanced Research in Computer and Communication Engineering 6, no. 1: 229–231. 10.17148/IJARCCE.2017.6143. [ DOI ] [ Google Scholar ] Penn State Extension . 2012. “Plant Disease Basics and Diagnosis.” Online. Petchiammal, A. Kiruba B., D. Murugan, and Arjunan P.. 2023. “Paddy Doctor: A Visual Image Dataset for Automated Paddy Disease Classification and Benchmarking.” 203–207. Quan, S. , Wang J., Jia Z., Yang M., and Xu Q.. 2023. “MS‐Net: A Novel Lightweight and Precise Model for Plant Disease Identification.” Frontiers in Plant Science 14: 1276728. 10.3389/fpls.2023.1276728. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Reda, M. , Suwwan R., Alkafri S., Rashed Y., and Shanableh T.. 2022. Agroaid: A Mobile App System for Visual Classification of Plant Species and Diseases Using Deep Learning and TensorFlow Lite. Vol. 9, 55. MDPI. [ Google Scholar ] Richter, D. J.

2025a. “PLDC‐80.” https://github.com/JDatPNW/PLDC‐80 . Richter, D. J.

2025b. “Developing a Strong Pre‐Trained Base Model for Plant Leaf Disease Classification.” PhD Thesis. Chonnam National University, Gwangju, South Korea; in Press. Richter, D. J.

2025c. “PLDC‐Net.” https://github.com/JDatPNW/PLDC‐Net . Richter, D. J. , Bappi M. I., Kolekar S. S., and Kim K.. 2025. “A Systematic Review of the Current State of Transfer Learning Accelerated CNN‐Based Plant Leaf Disease Classification.” IEEE Access 13: 116262–116303. 10.1109/ACCESS.2025.3584404. [ DOI ] [ Google Scholar ] Richter, D. J. , and Kim K.. 2025a. “Assessing the Performance of Domain‐Specific Models for Plant Leaf Disease Classification: A Comprehensive Benchmark of Transfer‐Learning on Open Datasets.” Scientific Reports 15, no. 1: 1–31. 10.1038/s41598-025-03235-w. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Richter, D. J. , and Kim K.. 2025b. “PLDC‐80: A Combined Augmented Hybrid Plant Leaf Disease Classification Benchmarking Dataset.” in Press. Salam, A. , Naznine M., Jahan N., Nahid E., Nahiduzzaman M., and Chowdhury M. E.. 2024. “Mulberry Leaf Disease Detection Using CNN‐Based Smart Android Application.” IEEE Access 12: 83575–83588. 10.1109/ACCESS.2024.3407153. [ DOI ] [ Google Scholar ] Sarkar, C. , Gupta D., Gupta U., and Hazarika B. B.. 2023. “Leaf Disease Detection Using Machine Learning and Deep Learning: Review and Challenges.” Applied Soft Computing 145: 110534. 10.1016/j.asoc.2023.110534. [ DOI ] [ Google Scholar ] Selvaraju, R. R. , Cogswell M., Das A., Vedantam R., Parikh D., and Batra D.. 2017. “Grad‐Cam: Visual Explanations From Deep Networks via Gradient‐Based Localization.” 618–626. Shah, S. K. , Kumbhar V., and Singh T.. 2024. “Grape ( Vitis vinifera ) Leaf Disease Detection and Classification Using Deep Learning Techniques: A Study on Real‐Time Grape Leaf Image Dataset in India.” International Journal of Engineering 37, no. 8: 1522–1533. 10.5829/ije.2024.37.08b.06. [ DOI ] [ Google Scholar ] Simon, M. , Rodner E., and Denzler J.. 2016. “Imagenet Pre‐Trained Models With Batch Normalization.” arXiv Preprint arXiv:1612.01452. 10.48550/arXiv.1612.01452. [ DOI ] Simonyan, K. , and Zisserman A.. 2014. “Very Deep Convolutional Networks for Large‐Scale Image Recognition.” arXiv Preprint arXiv:1409.1556. 10.48550/arXiv.1409.1556. [ DOI ] Singh, D. , Jain N., Jain P., Kayal P., Kumawat S., and Batra N.. 2020. PlantDoc: A Dataset for Visual Plant Disease Detection, 249–253. Association for Computing Machinery. [ Google Scholar ] Sultan, T. , Chowdhury M. S., Jahan N., et al. 2025. “LeafDNet: Transforming Leaf Disease Diagnosis Through Deep Transfer Learning.” Plant Direct 9, no. 2: e70047. 10.1002/pld3.70047. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Tan, M. , and Le Q.. 2019. Efficientnet: Rethinking Model Scaling for Convolutional Neural Networks, 6105–6114. PMLR. [ Google Scholar ] Tan, M. , and Le Q.. 2021. Efficientnetv2: Smaller Models and Faster Training, 10096–10106. PMLR. [ Google Scholar ] Thapa, R. , Zhang K., Snavely N., Belongie S., and Khan A.. 2020. “The Plant Pathology Challenge 2020 Data Set to Classify Foliar Disease of Apples.” Applications in Plant Sciences 8, no. 9: e11390. 10.1002/aps3.11390. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Thite, S. , Suryawanshi Y., Patil K., and Chumchu P.. 2024. “Sugarcane Leaf Dataset: A Dataset for Disease Detection and Classification for Machine Learning Applications.” Data in Brief 53: 110268. 10.1016/j.dib.2024.110268. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Winiarti, S. , and Pujiyanta A.. 2024. “Identification of Chili Plant Diseases Based on Leaves Using Hyperparameter Optimization Architecture Convolutional Neural Network.” International Journal of Advanced Computer Science and Applications 15, no. 11: 859–865. 10.14569/IJACSA.2024.0151185. [ DOI ] [ Google Scholar ] Xu, M. , Kim H., Yang J., et al. 2023. “Embracing Limited and Imperfect Training Datasets: Opportunities and Challenges in Plant Disease Recognition Using Deep Learning.” Frontiers in Plant Science 14: 1225409. 10.3389/fpls.2023.1225409. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Xu, M. , Park J. E., Lee J., Yang J., and Yoon S.. 2024. “Plant Disease Recognition Datasets in the Age of Deep Learning: Challenges and Opportunities.” Frontiers in Plant Science 15: 1452551. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. Articles from Plant Direct are provided here courtesy of Wiley ACTIONS View on publisher site PDF (5.3 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Reduced susceptibility to triclosan and proposed epidemiological cut-off values for disinfectants and heavy metals in Escherichia coli isolated from Thai tilapia farms: Implications for One Health antimicrobial resistance surveillance.#1039988
• Edible coatings and films in meat preservation: Materials, functional mechanisms, and technological challenges.#1039989
• Palm Vein Biometrics A Rapid Patient Identification Tool for Emergency and Trauma Care.#1039990
• Directions for Establishing a Korean National One Health Joint Plan of Action: Focusing on Zoonotic Disease Preparedness and Response.#1039991
• Intensive arable farming calls for resilient wild bee restoration amid severe abundance fluctuations.#1039992