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 Environ Monit Assess . 2026 Apr 11;198(5):447. doi: 10.1007/s10661-026-15297-y Search in PMC Search in PubMed View in NLM Catalog Add to search Deep learning-based classification of nitrate and nitrite concentrations from water samples using colorimetric test strip images Muhammad Roman Muhammad Roman 1 Department of Agricultural and Biosystems Engineering, South Dakota State University, Brookings, SD 57007 USA Find articles by Muhammad Roman 1 , Mazhar Sher Mazhar Sher 1 Department of Agricultural and Biosystems Engineering, South Dakota State University, Brookings, SD 57007 USA Find articles by Mazhar Sher 1, ✉ , Chamika Kuruppuarachchi Chamika Kuruppuarachchi 1 Department of Agricultural and Biosystems Engineering, South Dakota State University, Brookings, SD 57007 USA Find articles by Chamika Kuruppuarachchi 1 , Arshid Ali Arshid Ali 2 Department of Electrical Engineering and Computer Science, South Dakota State University, Brookings, SD 57007 USA Find articles by Arshid Ali 2 , Chulwoo Pack Chulwoo Pack 2 Department of Electrical Engineering and Computer Science, South Dakota State University, Brookings, SD 57007 USA Find articles by Chulwoo Pack 2 , Azlan Zahid Azlan Zahid 3 Department of Biological and Agricultural Engineering, Texas A&M AgriLife Research, Texas A&M University System, Dallas, TX 75252 USA Find articles by Azlan Zahid 3 , Ali Mirzakhani Nafchi Ali Mirzakhani Nafchi 1 Department of Agricultural and Biosystems Engineering, South Dakota State University, Brookings, SD 57007 USA 4 Department of Agronomy, Horticulture and Plant Science, College of Agriculture, Food and Environmental Sciences, South Dakota State University, Brookings, SD 57007 USA Find articles by Ali Mirzakhani Nafchi 1, 4 Author information Article notes Copyright and License information 1 Department of Agricultural and Biosystems Engineering, South Dakota State University, Brookings, SD 57007 USA 2 Department of Electrical Engineering and Computer Science, South Dakota State University, Brookings, SD 57007 USA 3 Department of Biological and Agricultural Engineering, Texas A&M AgriLife Research, Texas A&M University System, Dallas, TX 75252 USA 4 Department of Agronomy, Horticulture and Plant Science, College of Agriculture, Food and Environmental Sciences, South Dakota State University, Brookings, SD 57007 USA ✉ Corresponding author. Received 2025 Sep 1; Accepted 2026 Mar 31; Issue date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . PMC Copyright notice PMCID: PMC13070070 PMID: 41965491 Abstract Accurate monitoring of nitrate and nitrite concentrations in water is essential for sustainable agriculture, safeguarding public health, and protecting aquatic ecosystems from nutrient pollution. Traditional methods for detecting nitrate and nitrite in water samples are precise but costly, complex, and time-consuming, limiting their practicality for frequent on-site testing. This research proposes deep learning-based computer vision techniques to classify nitrate and nitrite concentrations using images of colorimetric test strips. An RGB IMX219 camera was used to acquire images of colorimetric test strips under standardized, controlled illumination conditions to ensure consistent image quality. A total of 1938 nitrate images and 1190 nitrite images were collected before augmentation. After preprocessing and training-only data augmentation, both classical machine learning baselines based on hand-crafted color and texture features and deep learning models—including a multilayer perceptron (MLP) and convolutional neural networks (AlexNet, VGG16, ResNet18, and GoogLeNet)—were trained and evaluated using an independent test set and stratified fivefold cross-validation. For nitrate classification, ResNet18 and GoogLeNet achieved near-perfect 100% test accuracy, with mean cross-validation accuracy of 99.97% ± 0.04%, substantially outperforming classical baseline models based on hand-crafted color and texture features, which achieved at most 83.5% test accuracy. For nitrite classification, GoogLeNet achieved the strongest overall performance, with a test accuracy of 97.48% and a fivefold cross-validation accuracy of 95.22% ± 1.17%, substantially outperforming the best classical baseline model, which achieved a maximum test accuracy of 83.19%. These results demonstrate that deep CNN-based feature learning provides a significant performance advantage over simpler methods under controlled imaging conditions, supporting the suitability of the proposed system for rapid, image-based water quality assessment and motivating future evaluation under broader real-world deployment scenarios. Keywords: Precision agriculture, Image classification, Water quality monitoring, Deep learning Introduction Water is a fundamental natural resource essential for the sustenance and survival of human life (Alizamir et al., 2025 ). In recent years, the accelerating pace of industrialization, urban development, and intensive agricultural activities has affected the water quality on a global scale (Akhtar et al., 2021 ; Saxena, 2025 ). Freshwater is increasingly threatened by human-caused pollutants, including chemical waste, heavy metals, pathogens, and organic matter, which tend to be discharged into river systems, lakes, and groundwater reservoirs without having been treated (Mishra, 2023 ; Mushtaq et al., 2020 ). These contaminants also pose a great threat to the environment and public health, particularly in regions with limited sources of clean water infrastructure. As a result, ensuring access to clean water is now one of the most urgent global health and environmental challenges of the twenty-first century (Elimelech, 2006 ; Levin et al., 2002 ). Therefore, the prediction and assessment of water quality constitute a critical task for ensuring environmental sustainability and safeguarding public health (Yan et al., 2022 ). Nitrate and nitrite are among the most critical chemical contaminants affecting water quality (Kensington et al., 2004 ). The primary sources of nitrate and nitrite contamination in water include agricultural runoff from excessive fertilizer use, as well as contributions from industrial waste, septic systems, and livestock manure. These compounds are leached into groundwater and are transported through surface water, due to their high solubility, thus making them prevalent in both rural and urban water supplies (Craswell, 2021 ). Prolonged exposure to these chemical compounds in drinking water can have serious health consequences, such as methemoglobinemia (known as “blue baby syndrome”) (Majumdar, 2003 ), increased risk of certain cancers (Picetti et al., 2022 ), and disruptions to the human endocrine and cardiovascular systems (Bryan & Loscalzo, 2017 ). These pollutants contribute to eutrophication, resulting in oxygen depletion and loss of aquatic biodiversity (Jan et al., 2022 ). Due to their impact on human health and the environment, continuous water quality monitoring is critical for regulatory compliance, public safety, and ecosystem protection. Standard methods for measuring nitrate and nitrite concentrations in water, such as ion chromatography, ultraviolet–visible (UV–Vis) spectrophotometry, and electrochemical methods, are recognized for their high quality and reliability in laboratory settings. Ion chromatography (IC) is a widely used technique for accurately separating and quantifying ions in water samples; however, it requires expensive instrumentation and skilled operation (Connolly & Paull, 2001 ). Spectrophotometry relies on the absorption characteristics of nitrate and nitrite compounds at specific wavelengths, providing reliable results but often involving complex reagents and sample preparation steps (Narayana & Sunil, 2009 ). Ion-selective electrodes and voltammetric techniques offer direct detection with reasonable sensitivity but are prone to interference from co-existing ions and often require frequent calibration (Revsbech et al., 2020 ). Despite their analytical accuracy, these methods are limited by high cost, complex sample preparation, dependence on laboratory infrastructure, and delayed processing times, which restrict their suitability for large-scale, real-time monitoring. Therefore, there is a clear need for rapid, affordable, and user-friendly methods for measuring nitrate and nitrite concentrations with high reliability and accuracy. Recent advances in artificial intelligence (AI), particularly in computer vision and deep learning, have opened new areas for efficient and automated water quality monitoring (Gunda et al., 2019 ; Ighalo et al., 2021 ; Yang et al., 2022 ). These techniques enable effective extraction of spatial features from visual data, significantly improving contaminant detection performance. Compared to traditional machine learning models, convolutional neural networks (CNNs) have demonstrated superior performance due to their ability to automatically learn hierarchical representations and effectively generalize across varied datasets (Esteva et al., 2017 ; LeCun et al., 2015 ). Specifically, GoogLeNet, ResNet18, and VGG16 models have been widely chosen for environmental and agricultural monitoring tasks because of their demonstrated robustness, effectiveness in feature extraction, and high classification accuracy (Wu et al., 2025 ; Yang et al., 2021 ; Yuesheng et al., 2021 ). Unlike traditional methods, AI-based systems can be deployed in low-cost, portable setups, enabling rapid on-site assessments with minimal human intervention. Vision-based analysis of colorimetric test strips has emerged as a promising alternative for nitrate measurement using camera or smartphone-assisted approaches. However, many existing methods rely on manual RGB feature extraction, calibration curves, and uncontrolled imaging conditions, which can limit robustness and repeatability. Prior studies have reported that smartphone-based quantification of nitrate test strips does not consistently outperform visual interpretation, highlighting the influence of imaging conditions and user-dependent variability (Topping & Kolok, 2021 ). Moreover, vision-based systems targeting fine-grained, multi-class classification of nitrate and nitrite concentrations remain limited. These challenges motivate the need for learning-based approaches that can automatically capture subtle spatial and color patterns under standardized imaging conditions. In this study, we propose a computer vision-based pipeline that leverages both classical and deep learning models to detect and classify nitrate and nitrite concentrations from the water test strip images. Classical baseline models, including logistic regression (LR), linear support vector machine (SVM), and Naïve Bayes (NB), are incorporated to provide interpretable reference performance and to quantify the benefits of deep feature learning. In parallel, several deep learning architectures are evaluated, including a multilayer perceptron (MLP) and several convolutional neural networks (CNNs) such as AlexNet (Pu et al., 2019 ), VGG16 (Al-Sudani & Al-Suhail, 2024 ), ResNet18 (Malathi et al., 2023 ), and GoogLeNet (Sumaya et al., 2024 ), each selected for their proven effectiveness in image classification tasks. Using a standardized image dataset, we evaluated each model’s learning capability, classification accuracy, and computational efficiency to assess their suitability for real-time, scalable deployment in water quality monitoring applications. Key contributions include the application of multiple deep learning models for multi-class classification of nitrate and nitrite concentrations and a comprehensive performance analysis to identify the most effective models for future integration into AI-powered water monitoring systems. Methodology The overall architecture of the proposed system comprises a streamlined pipeline for classifying nitrate and nitrite concentrations from test strip dataset images using deep learning techniques. The system is composed of three main stages: dataset preparation, model training, and evaluation. In the first stage, the labeled dataset of nitrate and nitrite is prepared and divided into 80:10:10 ratios for training, validation, and testing. The dataset was preprocessed and resized to a standard size. The trained dataset was augmented. In the second step, latent features were extracted from the training dataset using five deep learning models. Model training and validation were performed using Google Colab, leveraging GPU acceleration to reduce computation time and optimize model performance. Each model received the same input data to ensure a fair performance comparison. Finally, in the evaluation stage, the models were assessed using evaluation matrices to visualize their performance across different classes. The overall system architecture is shown in Fig. 1 . Fig. 1. Open in a new tab System Architecture flow diagram of nitrate and nitrite classification Dataset Commercially available Hach water quality test strips for nitrate and nitrite were used to collect the image dataset. The test strips were immersed in the water samples with known concentrations. After color development, the test strip image was captured with an IMX219 camera inside a custom-designed, 3D-printed, enclosed imaging box under controlled lighting. The enclosed setup ensured a fixed camera position and illumination geometry, minimizing the influence of ambient lighting, shadows, and background variations during image acquisition. Each image is stored according to the corresponding nitrate and nitrite concentration, resulting in a well-structured dataset suitable for deep learning applications. The specific concentrations used for sample preparation are detailed in Table 1 . To enhance computational efficiency, the images were preprocessed to reduce noise using a median filter and uniformly resized to 64 × 64 pixels. This dataset was created by the Biosensors lab SDSU (Muhammad Roman et al., 2026 ) . Table 1. Nitrate and nitrite sample concentrations Nitrate samples (PPM) Nitrite samples (PPM) 0, 1, 2, 5, 10, 15, 20, 30, 40, 50 0.15, 0.3, 1.0, 1.5, 3.0 Open in a new tab a. Nitrate-nitrite dataset The initial image datasets for both nitrate and nitrite classification contained approximately 200 images per class. Each image was captured using a new test strip with a known concentration of nitrate and nitrite. The dataset was divided into training, validation, and test sets using an 80/10/10 split on a per-class basis to preserve class balance across all concentrations. Data augmentation was applied only to the training set, while validation and test sets remained unchanged to ensure unbiased performance evaluation. Augmentation techniques included random rotations (± 90°), horizontal flipping, and controlled brightness variation (± 10%) to simulate minor illumination changes, while the imaging system itself operated in an enclosed environment to minimize external lighting effects. The distribution of training and testing images per class, along with total counts, is presented in Table 2 for nitrate and Table 3 for nitrite. Representative sample images from each class are shown in Figs. 2 and 3 , respectively. Table 2. Nitrate dataset images before and after augmentation Class (ppm) Train (before Aug) Train (after Aug) Val Test Total Nitrate 0 94 376 12 12 400 1 180 720 23 23 766 2 171 684 22 22 728 5 153 612 19 19 650 10 177 708 22 22 752 15 137 548 17 17 582 20 163 652 20 20 692 30 155 620 19 19 658 40 156 624 19 19 662 50 164 656 21 21 698 Total 1550 6200 194 194 6588 Open in a new tab Table 3. Nitrite dataset images before and after augmentation Class (ppm) Train (before Aug) Train (after Aug) Val Test Total Nitrite 0 132 528 16 16 560 0.15 167 668 21 21 710 0.3 168 672 21 21 714 1 166 664 21 21 706 1.5 164 656 20 20 696 3 155 620 20 20 660 Total 952 3808 119 119 4046 Open in a new tab Fig. 2. Open in a new tab Random 1 sample from each nitrate class Fig. 3. Open in a new tab Random 1 sample from each nitrite class Deep learning architecture This study utilizes several deep learning models to classify nitrate and nitrite concentrations from colorimetric strip images. The selected models include traditional architectures (MLP, AlexNet, VGG16) and more advanced designs (ResNet18, GoogLeNet), enabling a comparative analysis of performance across different levels of complexity and feature learning capabilities. a) Multilayer perceptron The Multilayer Perceptron is a feedforward neural network composed of dense, fully connected layers, as shown in Fig. 4 . It accepts vectorized input and relies on non-linear activations to learn feature representations. MLPs are often used as baselines due to their simplicity and fast training, but they lack spatial context awareness inherent to image data (Foody, 2004 ). Fig. 4. Open in a new tab Multilayer perceptron network model architecture b) AlexNet AlexNet is one of the first deep CNNs to demonstrate the potential of deep learning for image recognition tasks. The network architecture (Lorenzo et al., 2020 ) is shown in Fig. 5 . It consists of five convolutional layers followed by three fully connected layers and introduces the use of ReLU activations and dropout regularization. Although pioneering, AlexNet’s architecture includes large filters and pooling steps that can limit its performance on low-resolution images (Chen et al., 2022 ). Fig. 5. Open in a new tab AlexNet model architecture c) VGG16 VGG16 is a deep CNN architecture that uses a stack of small 3 × 3 convolutional filters across 16 weight layers. Its uniform and simple structure has made it widely adopted for visual recognition benchmarks. However, the large number of parameters in VGG16 can lead to high memory consumption and overfitting when used on smaller datasets (Simonyan & Zisserman, 2014 ). The network architecture (Shazia et al., 2021 ) is shown in Fig. 6 . Fig. 6. Open in a new tab VGG16 network architecture d) ResNet18 ResNet18 introduced the concept of residual learning by incorporating skip connections that bypass one or more layers. These identity mappings improve gradient flow during backpropagation, enabling the training of deeper networks with fewer convergence issues. ResNet18 is a widely used balanced architecture known for its efficiency and strong generalization (He et al., 2016 ). Figure 7 highlights the core components of the ResNet18 Network (MohammadiNasab, 2023 ). Fig. 7. Open in a new tab ResNet18 model architecture e) GoogLeNet GoogLeNet is a CNN architecture that utilizes inception modules, combinations of 1 × 1, 3 × 3, and 5 × 5 convolutions executed in parallel, to efficiently capture multiscale spatial information. The inclusion of 1 × 1 convolutions reduces dimensionality, making the network both deep and computationally efficient. It also incorporates auxiliary classifiers to stabilize training and reduce vanishing gradients (Abd El-Maksoud et al., 2021 ). GoogLeNet system architecture (Alkan et al., 2021 ) is shown in Fig. 8 . Fig. 8. Open in a new tab Block diagram of the GoogLeNet architecture with inception modules used for classification Classical machine learning baseline models In addition to deep learning architectures, classical machine learning models were implemented to provide baseline performance comparisons and to quantify the benefit of deep feature learning. These models rely on hand-crafted image features rather than end-to-end representation learning and are commonly used in traditional colorimetric and image-based analysis tasks. Three simple yet representative feature sets were extracted from the colorimetric strip images: mean HSV color features to capture global color intensity, hue histograms to represent color distribution variations across concentrations, and local binary pattern (LBP) histograms to describe basic texture information within the reaction regions. Using these features, three baseline classifiers were evaluated: logistic regression (Arief Subchan & Andayani, 2021 ), linear support vector machine (SVM) (Ebrahimi et al., 2017 ), and Gaussian Naïve Bayes (Agarwal et al., 2018 ). All baseline models were trained and evaluated using the same train/validation/test splits as the deep learning models to ensure a fair comparison. These classical baselines provide interpretable reference models and highlight the limitations of hand-crafted features when compared to deep convolutional neural networks. Training Each model was trained using a categorical cross-entropy loss function optimized with the Adam optimizer. The initial learning rate was set to 0.001 for MLP and ResNet18, and 0.0005 for AlexNet, VGG16, and GoogLeNet, with a batch size of 32. Training was conducted for up to 50 epochs, with early stopping based on validation loss to prevent overfitting. To ensure stable convergence, a StepLR learning rate scheduler was employed with a decay factor of 0.1 applied at fixed intervals during training. Model performance was monitored using a dedicated validation set, while the test set was used only for final evaluation. All experiments were conducted using fixed random seeds and deterministic training settings to ensure reproducibility. The MLP model consisted of fully connected layers with ReLU activation functions. Due to its straightforward structure, the MLP served as a baseline for comparative performance analysis. AlexNet was trained with its original architecture of five convolutional layers followed by three fully connected layers. Dropout regularization was applied after the fully connected layers to mitigate overfitting. The VGG16 model fine-tunes its original deep convolutional architecture. Fine-tuning involved training the last convolutional block along with fully connected layers to adapt the model specifically to the nitrate and nitrite image dataset. The ResNet18 model leveraged residual connections to maintain effective gradient flow during training. Fine-tuning was performed by training all residual blocks with emphasis on maintaining the integrity of skip connections. The GoogLeNet architecture incorporated inception modules to capture multiscale features effectively. Auxiliary classifiers within GoogLeNet assist in stabilizing the training process and improving convergence speed, leading to superior classification accuracy. In contrast to deep learning models, classical machine learning baselines do not require iterative training over epochs. Instead, feature extraction was performed once per image, followed by direct classifier fitting on the training set. Feature normalization was applied prior to training, and model evaluation was performed using the same validation and test sets as the deep learning experiments to maintain consistency. Performance evaluation Performance metrics are crucial for assessing the effectiveness of deep learning classification models, providing quantitative insights into a model’s ability to accurately predict class labels. A fundamental tool for this assessment is the confusion matrix, which compares predicted labels with actual labels and consists of four components: true positive (TP), true negative (TN), false positive (FP), and false negative (FN). Based on the confusion matrix, performance metrics such as accuracy, precision, recall, and F1-score are computed to evaluate classification performance (Eq. 1 – 4 ). Accuracy reflects the proportion of correct predictions out of the total predictions made (Anderson, 2018 ; Hernandez-Guedes et al., 2022 ; Khalid, 2024 ). In addition, model performance is evaluated using training and testing accuracy and loss curves (Eq. 5 ). A fivefold stratified cross-validation strategy was employed on the training set to ensure robust performance assessment given the relatively limited dataset size. For each fold, models were trained on four subsets and evaluated on the remaining subset, and the final performance was reported as the mean accuracy ± standard deviation across all folds. This approach provides a reliable estimate of model stability and generalization. Training and testing accuracy and loss curves provide insight into model convergence and generalization behavior. Divergence between training and testing performance typically indicates overfitting or unstable learning. A c c u r a c y = T P + T N T P + T N + F P + F N 1 P r e c i s i o n = TP T P + F P 2 R e c a l l = TP T P + F N 3 F 1 - s c o r e = 2 p r e c i s i o n ∗ r e c a l l p r e c i s i o n + r e c a l l 4 c r o s s - e n t r o p y l o s s = - ∑ i = 1 N ∑ c = 1 C y i , c . log y ^ i , c 5 Results and discussion Nitrate classification The performance of each model was evaluated on the independent test set using confusion matrices, along with accuracy, precision, recall, and F1-score metrics. These metrics provide a comprehensive assessment of both overall classification accuracy and class-wise behavior, highlighting patterns of misclassification and model robustness. The confusion matrices and detailed precision, recall, and F1-scores for nitrate classification are presented in Fig. 9 . Fig. 9. Open in a new tab Confusion matrices for nitrate classification using different models: a classical baseline classifiers (logistic regression, linear support vector machine, and Naïve Bayes), b multilayer perceptron (MLP), c AlexNet, d VGG16, e ResNet18, and f GoogLeNet. Each confusion matrix shows the model’s ability to distinguish among ten nitrate concentration classes (0, 1, 2, 5, 10, 15, 20, 30, 40, and 50 ppm). Strong diagonal dominance indicates accurate classification, with ResNet18 and GoogLeNet demonstrating near-perfect performance across all concentrations. g Precision, recall, and F1-score values for each class and model Among the deep learning models, the MLP exhibited limited performance, particularly at higher nitrate concentrations. While lower concentration classes (0, 1, 2, 5, and 10 ppm) were classified reliably, substantial confusion occurred at 30, 40, and 50 ppm. Notably, the 40 ppm class was not correctly predicted at all, resulting in zero recall and F1-score, while the 30 ppm class achieved a recall of only 0.11 and an F1-score of 0.13. These results highlight the inherent limitation of MLP architectures for image-based tasks, as they lack spatial feature extraction capabilities. AlexNet demonstrated a significant improvement over MLP, achieving an overall test accuracy of 96.9%. Most nitrate classes were classified with high precision and recall; however, moderate confusion was observed between adjacent concentrations, particularly at 30 ppm and 50 ppm, where F1-scores dropped to 0.86 and 0.90, respectively. Despite these errors, AlexNet showed strong generalization compared to simpler architectures. VGG16 achieved a test accuracy of 89.2%, but exhibited inconsistent performance across mid-range nitrate concentrations. Although perfect classification was observed for several low and high concentration classes, the 30 ppm and 40 ppm classes suffered from notable misclassification, with F1-scores of 0.62 and 0.17, respectively. This variability may be attributed to VGG16’s large parameter count, which can lead to suboptimal generalization when training data are limited. In contrast, ResNet18 and GoogLeNet achieved perfect classification performance, each reaching 100% test accuracy, with precision, recall, and F1-scores of 1.00 across all nitrate classes. The residual connections in ResNet18 and the multiscale feature extraction enabled by GoogLeNet’s inception modules contributed to stable optimization and exceptional generalization. Stratified fivefold cross-validation further confirmed this robustness, with mean accuracies exceeding 99.9% for both architectures. In addition to deep learning models, classical machine learning baselines based on hand-crafted features were evaluated to contextualize the benefit of CNN-based approaches. The logistic regression baseline achieved a test accuracy of 83.5%, performing reasonably well at low concentrations but struggling to discriminate higher nitrate levels. The linear SVM model performed poorly, with a test accuracy of 56.7%, while the Naïve Bayes baseline failed to generalize effectively, achieving only 15.5% accuracy. These results demonstrate that color and texture features alone are insufficient to capture the subtle spatial variations present in colorimetric test strip images. Overall, the results clearly indicate that modern CNN architectures, particularly ResNet18 and GoogLeNet, substantially outperform both simpler deep models and classical machine learning baselines, making them well suited for accurate and real-time nitrate concentration classification under standardized imaging conditions. Nitrite classification The performance of the proposed models for nitrite classification was evaluated using confusion matrices and class-wise precision, recall, and F1-score metrics on the independent test set. The confusion matrices for nitrite classification are presented in Fig. 10 , providing detailed insight into class-level prediction behavior and model robustness. Fig. 10. Open in a new tab Confusion matrices for nitrite classification using different models: a classical baseline classifiers (logistic regression, linear support vector machine, and Naïve Bayes), b multilayer perceptron (MLP), c AlexNet, d VGG16, e ResNet18, and f GoogLeNet. Each confusion matrix shows the model’s ability to distinguish among six nitrite concentration classes (0, 0.15, 0.3, 1, 1.5, and 3 ppm). Strong diagonal dominance indicates accurate classification, with ResNet18 and GoogLeNet demonstrating near-perfect performance across all concentrations. g Precision, recall, and F1-score values for each class and model Among the deep learning models, GoogLeNet achieved the best overall performance, with a test accuracy of 97.5% and consistently high precision and recall across all nitrite concentrations. Minor misclassifications were observed only between the 0.15 ppm and 0.3 ppm classes, where subtle color variations and overlapping visual characteristics likely increased classification difficulty. Nevertheless, GoogLeNet maintained macro and weighted F1-scores above 0.97, demonstrating excellent generalization capability. ResNet18 also showed strong performance, achieving a test accuracy of 91.6%. Most nitrite classes were classified reliably; however, moderate confusion occurred between the 0.15 ppm and 0.3 ppm classes, resulting in slightly reduced recall values (0.71 and 0.81, respectively). Despite these errors, ResNet18 maintained balanced performance across all concentrations, supported by a macro F1-score of 0.92. In contrast, MLP, AlexNet, and VGG16 showed comparatively lower performance, particularly for low-concentration nitrite classes. The MLP model failed to correctly identify the 0.15 ppm class, resulting in zero recall and a reduced test accuracy of 82.35%, highlighting the limitations of fully connected architectures for spatially complex image data. AlexNet and VGG16 achieved test accuracies of 87.39% and 89.92%, respectively, but both models exhibited noticeable confusion between visually similar low-level concentrations, especially 0.15 ppm and 0.3 ppm. To further contextualize these results, classical machine learning baseline models were also evaluated using hand-crafted visual features. Logistic regression achieved a test accuracy of 83.19%, demonstrating that color information alone provides a reasonable baseline for nitrite estimation. However, this approach struggled with mid-range concentrations, where overlapping color distributions reduced class separability. A linear SVM yielded a lower test accuracy of 61.34%, indicating insufficient discriminative power for complex multi-class separation. The Naïve Bayes model performed poorly, achieving a test accuracy of only 21.85%, confirming that texture features alone are not suitable for nitrite strip classification. Overall, these findings clearly demonstrate that deep CNN architectures significantly outperform classical hand-crafted feature-based models for nitrite classification. The superior performance of GoogLeNet and ResNet18 highlights the importance of hierarchical feature learning for capturing subtle color and spatial variations, making these architectures more suitable for reliable, high-precision nitrite monitoring under controlled imaging conditions. Model training and validation performance The training and validation performance of the five deep learning models (MLP, AlexNet, VGG16, ResNet18, and GoogLeNet) for nitrate and nitrite classification was systematically analyzed using accuracy and loss curves, as shown in Figs. 11 and 12 . These curves provide insight into convergence behavior, learning stability, and potential overfitting across epochs. Fig. 11. Open in a new tab Training and testing accuracy and loss curves for nitrate classification using five deep learning models: MLP, AlexNet, VGG16, ResNet18, and GoogLeNet Fig. 12. Open in a new tab Training and testing accuracy and loss curves for nitrite classification using five deep learning models: MLP, AlexNet, VGG16, ResNet18, and GoogLeNet For nitrate classification (Fig. 11 ), all models exhibited a general increase in training accuracy accompanied by a decrease in training loss during the initial epochs, indicating effective feature learning. However, clear differences were observed in validation behavior. GoogLeNet and ResNet18 demonstrated highly stable validation accuracy curves that closely followed their training accuracy, along with consistently decreasing validation loss. This strong alignment between training and validation trends indicates excellent generalization and minimal overfitting. AlexNet showed moderate fluctuations in validation accuracy and loss across epochs, suggesting sensitivity to intra-class variations among nitrate concentrations. VGG16 exhibited a slower convergence rate with higher validation loss in early epochs, but its performance gradually improved as training progressed, reflecting the impact of its deeper architecture on feature extraction. The MLP model showed the slowest convergence and the largest gap between training and validation accuracy, highlighting its limited ability to capture spatial and chromatic patterns from image data. Overall, GoogLeNet and ResNet18 achieved the most consistent and reliable training–validation behavior for nitrate classification. For nitrite classification (Fig. 12 ), all models showed an overall increase in training accuracy and a decrease in training loss across epochs, indicating effective learning. However, greater variability was observed in validation behavior compared to nitrate classification, reflecting the increased complexity of nitrite strip patterns. GoogLeNet demonstrated the most stable performance, with closely aligned training and validation accuracy curves and consistently decreasing validation loss, indicating strong generalization. ResNet18 also achieved high accuracy with relatively smooth convergence, despite minor early fluctuations in validation performance. In contrast, VGG16 exhibited moderate instability during early epochs, with fluctuating validation accuracy and loss before stabilizing. AlexNet showed larger oscillations in validation accuracy and occasional validation loss spikes, suggesting sensitivity to data variability. The MLP displayed the slowest convergence and the largest gap between training and validation accuracy, highlighting its limited capacity to capture complex features. Overall, GoogLeNet and ResNet18 provided the most reliable performance for nitrite classification. Model comparison Table 4 presents a comprehensive comparison of deep learning models and classical baseline approaches for both nitrate and nitrite classification. The comparison includes training, validation, and test accuracies, test loss, cross-validation performance, and computational efficiency, enabling a clear assessment of performance, generalization, and deployment feasibility across model families. Table 4. Performance comparison of deep learning and classical baseline models Architecture Train Acc Val Acc Test Acc Test Loss 5-fold CV Acc (mean ± std) Training time (s) Inference time (ms/sample) Nitrate Logistic regression 0.8098 0.8247 0.8351 — — < 1 0.0039 Linear SVM 0.5603 0.5361 0.5670 — — < 1 0.0037 Naïve Bayes 0.3455 0.1649 0.1546 — — < 1 0.0072 MLP 0.8545 0.7371 0.7268 0.5949 0.7392 ± 0.0484 75.44 0.0312 AlexNet 0.9316 0.9124 0.9691 0.171 0.9437 ± 0.0068 99.3 0.1934 VGG16 0.8571 0.8402 0.8711 0.3161 0.8484 ± 0.0343 652.22 0.5924 ResNet18 0.9913 1 1 0.0022 0.9997 ± 0.0004 128.51 1.8655 GoogLeNet 0.9979 1 1 0.0006 0.9997 ± 0.0004 187.93 0.843 Nitrite Logistic Regression 0.8382 0.8487 0.8319 — — < 1 0.0079 Linear SVM 0.5943 0.6555 0.6134 — — 2.03 0.0055 Naïve Bayes 0.4404 0.2185 0.2185 — — < 1 0.0089 MLP 0.8474 0.8319 0.8235 0.2498 0.8346 ± 0.0117 69.23 0.0216 AlexNet 0.8729 0.8824 0.8739 0.2074 0.8742 ± 0.0132 98.92 0.1172 VGG16 0.8427 0.8908 0.8992 0.2170 0.8600 ± 0.0213 326.26 0.5547 ResNet18 0.9070 0.9496 0.9160 0.1480 0.8868 ± 0.0275 122.25 0.2762 GoogLeNet 0.9603 1 0.9748 0.0870 0.9522 ± 0.0117 141.67 0.4671 Open in a new tab a) Deep learning models For nitrate classification, deep learning models consistently outperformed classical baselines across all evaluation stages. ResNet18 and GoogLeNet demonstrated the strongest and most stable performance, achieving near-perfect training accuracies (0.9913 and 0.9979), validation accuracies of 1.0000, and test accuracies of 1.0000, with very low-test losses (0.0022 and 0.0006, respectively). The close alignment between training, validation, test, and cross-validation results (0.9997 ± 0.0004 for both models) indicates excellent generalization and minimal overfitting. These results highlight the effectiveness of deeper architectures with residual connections and inception modules in capturing subtle chromatic and spatial variations in nitrate test strip images. Among the remaining deep models, AlexNet achieved high performance with a test accuracy of 0.9691, supported by strong training (0.9316) and validation (0.9124) accuracies, though minor confusion between adjacent concentration classes remained. VGG16 showed moderate performance (test accuracy 0.8711) with a noticeable reduction in cross-validation accuracy (0.8484 ± 0.0343) and substantially higher training time, suggesting increased sensitivity to harder or ambiguous samples. The MLP baseline exhibited the weakest deep learning performance, with a clear drop from training accuracy (0.8545) to test accuracy (0.7268), underscoring the limitations of fully connected architectures for image-based classification tasks. For nitrite classification, a similar performance hierarchy was observed. GoogLeNet again demonstrated the strongest results, achieving training, validation, and test accuracies of 0.9603, 1.0000, and 0.9748, respectively, with a low-test loss (0.0870) and stable cross-validation performance (0.9522 ± 0.0117). ResNet18 also showed strong generalization (test accuracy, 0.9160; CV, 0.8868 ± 0.0275), though performance was slightly reduced compared to nitrate classification, reflecting the increased difficulty of distinguishing low nitrite concentration levels. AlexNet and VGG16 demonstrated moderate and consistent performance, while MLP again showed reduced robustness, particularly in the presence of subtle inter-class variations. b) Classical baseline models Classical baseline methods provided a useful reference for quantifying the benefit of deep feature learning. For nitrate classification, logistic regression (LR) achieved the strongest baseline performance, with training, validation, and test accuracies of 0.8098, 0.8247, and 0.8351, respectively. In contrast, linear SVM and Naïve Bayes showed substantially lower performance, with test accuracies of 0.5670 and 0.1546, indicating limited discriminative capability. A similar trend was observed for nitrite classification, where logistic regression again produced the best baseline results (train, 0.8382; val, 0.8487; test, 0.8319). Linear SVM and Naïve Bayes failed to generalize effectively, particularly at lower concentration levels. Across both datasets, classical baselines exhibited larger discrepancies between training and test performance and higher inter-class confusion, highlighting their limited ability to model subtle color and texture variations compared to deep learning approaches. c) Computational efficiency and practical implications From a computational perspective, GoogLeNet and ResNet18 offered a favorable balance between accuracy and efficiency, achieving strong generalization with inference times below 1 ms per sample, making them suitable for real-time, on-site deployment. Although VGG16 achieved reasonable classification performance, its significantly higher training and inference costs reduced its practical appeal. Classical baseline models were computationally lightweight but did not achieve sufficient accuracy for reliable deployment. Overall, the results in Table 4 demonstrate that modern deep convolutional architectures—particularly GoogLeNet and ResNet18—provide substantial and consistent performance gains over classical baseline classifiers (LR, SVM, NB). The strong agreement between training, validation, test, and cross-validation results supports their suitability for robust, automated, vision-based nitrate and nitrite monitoring systems under controlled imaging conditions. Evaluation using independent field samples To further evaluate the practical applicability of the proposed system, independent real water samples were collected from field locations and analyzed in the laboratory using an AQ2 discrete analyzer at South Dakota State University (SDSU) to obtain reference nitrate and nitrite concentrations. The same samples were subsequently tested using the colorimetric strips and classified by the trained GoogLeNet model. As shown in Table 5 , the predicted concentration classes generally aligned with the corresponding laboratory-measured values, with most samples correctly assigned to the nearest discrete concentration level. Minor deviations were primarily observed near class boundaries, as expected given the categorical nature of colorimetric strip readings and the inherent variability in low-concentration ranges. These results demonstrate that the proposed vision-based approach can provide reliable categorical estimates on real water samples under standardized imaging conditions, supporting its potential for rapid, on-site screening applications. Further evaluation under diverse environmental conditions is planned to assess robustness beyond controlled laboratory settings. Table 5. AQ2 laboratory measurements and corresponding GoogLeNet predictions for nitrate and nitrite concentrations in real field water samples Sample name Actual nitrate ppm GoogLeNet prediction ppm Actual nitrite ppm GoogLeNet prediction ppm Sample 1 (B1) 5.2445 5 0.05275 0 Sample 2 (B2) 1.4935 2 0.0047 0 Sample 3 (AUTO NORTH) 5.0275 5 0.3257 0.3 Sample 4 (AUTO SOUTH) 11.281 10 0.02265 0 Sample 5 (B3) 3.4655 5 0.02675 0 Sample 6 (B4) 0.44 0 0.02005 0 Sample 7 (B1) 2.631 2 0.0063 0 Sample 8 (B2) 0.6665 0 0.05735 0 Sample 9 (AUTO NORTH) 3.077 5 0.0076 0 Sample 10 (B3) 2.1795 2 0.008 0 Sample 11 (B4) 0.183 0 0.03165 0 Sample 12 (B1) 1.0215 1 0.0062 0 Sample 13 (AUTO NORTH) 1.8695 2 0.0081 0 Sample 14 (AUTO SOUTH) 1.551 2 0.0062 0 Sample 15 (B3) 1.7325 2 0.0157 0 Sample 16 (B4) 0.287 0 0.0333 0 Open in a new tab Conclusion This study presents a computer vision–based approach for classifying nitrate and nitrite concentrations using colorimetric test strip images. Multiple deep learning models, including MLP, VGG16, AlexNet, ResNet18, and GoogLeNet, were evaluated using the same datasets under consistent preprocessing and training conditions. Among these, ResNet18 and GoogLeNet consistently achieved the strongest performance, exhibiting near-perfect classification accuracy for nitrate and high accuracy (up to 97.48%) for nitrite classification under standardized imaging conditions. While models such as MLP, VGG16, and AlexNet served as comparative baselines, AlexNet notably underperformed due to its earlier-generation CNN architecture and limited capacity for multiscale feature representation, which hindered its ability to capture fine-grained spatial variations in the test strip images. Performance was generally lower for nitrite classification, likely due to greater intra-class variability, which demands stronger feature representation. Overall, the findings highlight the effectiveness of modern, deeper CNN architectures for developing accurate and reproducible vision-based systems for water quality monitoring under controlled imaging conditions. The current evaluation was conducted under controlled imaging conditions and does not explicitly assess robustness to environmental variations such as illumination changes, noise, strip orientation, or partial occlusion. The proposed solution shows strong potential for real-time processing and on-site deployment under standardized imaging conditions, particularly in resource-constrained settings. Future work will focus on optimizing and deploying the proposed models on low-power edge devices (e.g., NVIDIA Jetson–class platforms or Raspberry Pi–based systems) capable of real-time, on-device inference within an enclosed imaging setup, as well as evaluating robustness under broader environmental conditions to support practical and accessible water quality monitoring. Acknowledgements The authors would like to acknowledge the support of John Maursetter from the South Dakota Water Resources Institute. Author contribution Muhammad Roman: Methodology, Formal analysis, Investigation, Data curation, Writing—original draft. Chamika Kuruppuarachchi: Investigation, Data curation. Arshid Ali: Investigation, Data curation. Chulwoo Pack: Supervision and Writing—original draft. Azlan Zahid: Supervision and Writing—original draft. Ali Mirzakhani Nafchi: Supervision, Writing—original draft. Mazhar Sher: Project administration, Conceptualization, Resources, Supervision, Writing—original draft. Funding This research was funded by (1) South Dakota Nutrient Research Council (3S6184), (2) South Dakota Corn Utilization Council (3X6134), (3) The Midwest Dairy Council (3X6077), (4) The National Science Foundation RII Track-2 FEC: Advancement of Microbial Technologies for Sustainable Agriculture (AMiTSA) Award # 2316502, (5) the Hatch Project (3AH788), (6) Multi Hatch Project (NC1194) funded by USDA NIFA through South Dakota Agricultural Experimental Station at South Dakota State University, (7) The U.S. Department of Agriculture's National Institute of Food and Agriculture Women and Minorities in Science, Technology, Engineering, and Mathematics Fields Program, RASISE: Robotics, AI, and Sensors Education in K-14 Agriculture & STEM Grant No.2024-38503-43665, and (8) The United States National Science Foundation (NSF) (Grant OISE-2434687). Data availability Data will be provided on request. Declarations Ethical approval All authors have read, understood, and complied with the statement on “Ethical Responsibilities of Authors” found in the instructions to authors. Consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Abd El-Maksoud, A. J., Ebbed, M., Khalil, A. H., & Mostafa, H. (2021). Power efficient design of high-performance convolutional neural networks hardware accelerator on FPGA: A case study with GoogLeNet. IEEE Access, 9 , 151897–151911. [ Google Scholar ] Agarwal, S., Bhangale, N., Dhanure, K., Gavhane, S., Chakkarwar, V. A, & Nagori, M. B. (2018). Application of colorimetry to determine soil fertility through naive bayes classification algorithm. In 2018 9th International Conference on Computing, Communication and Networking Technologies (ICCCNT) (pp. 1–6). 10.1109/ICCCNT.2018.8494113 Akhtar, N., Syakir Ishak, M. I., Bhawani, S. A., & Umar, K. (2021). Various natural and anthropogenic factors responsible for water quality degradation: A review. Water, 13 (19), Article 2660. [ Google Scholar ] Alizamir, M., Moradveisi, K., Ahmed, K. O., Bahrami, J., Kim, S., & Heddam, S. (2025). An efficient data fusion model based on Bayesian model averaging for robust water quality prediction using deep learning strategies. Expert Systems with Applications, 261 , Article 125499. [ Google Scholar ] Alkan, A., Abdullah, M. U., Abdullah, H. O., Assaf, M., & Zhou, H. (2021). A smart agricultural application: Automated detection of diseases in vine leaves usinghybrid deep learning. Turkish Journal of Agriculture and Forestry, 45 (6), 717–729. [ Google Scholar ] Al-Sudani, A., & Al-Suhail, G. A. (2024). A comparative study of VGG16 and VGG19 for oil spill detection in coastal imagery. In 2024 1st International Conference on Emerging Technologies for Dependable Internet of Things (ICETI) (pp. 1–8). 10.1109/ICETI63946.2024.10777149 Anderson, C. B. (2018). The CCB-ID approach to tree species mapping with airborne imaging spectroscopy. PeerJ, 6 , Article e5666. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Arief Subchan, M., & Andayani, N. N. (2021, November). Classification of maize genotype using logistic regression. In IOP Conference Series: Earth and Environmental Science (Vol. 911, No. 1, p. 012017). IOP Publishing. 10.1088/1755-1315/911/1/012017 Bryan, N. S., & Loscalzo, J. (Eds.). (2017). Nitrite and nitrate in human health and disease . 10.1007/978-3-319-46189-2 Chen, H.-C., Widodo, A. M., Wisnujati, A., Rahaman, M., Lin, J.-W., Chen, L., & Weng, C.-E. (2022). AlexNet convolutional neural network for disease detection and classification of tomato leaf. Electronics, 11 (6), Article 951. [ Google Scholar ] Connolly, D., & Paull, B. (2001). Rapid determination of nitrate and nitrite in drinking water samples using ion-interaction liquid chromatography. Analytica Chimica Acta, 441 (1), 53–62. [ Google Scholar ] Craswell, E. (2021). Fertilizers and nitrate pollution of surface and ground water: An increasingly pervasive global problem. SN Applied Sciences, 3 (4), Article 518. [ Google Scholar ] Ebrahimi, M., Khoshtaghaza, M. H., Minaei, S., & Jamshidi, B. (2017). Vision-based pest detection based on SVM classification method. Computers and Electronics in Agriculture, 137 , 52–58. [ Google Scholar ] Elimelech, M. (2006). The global challenge for adequate and safe water. Journal of Water Supply: Research and Technology, 55 (1), 3–10. [ Google Scholar ] Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., & Thrun, S. (2017). Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542 (7639), 115–118. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Foody, G. (2004). Supervised image classification by MLP and RBF neural networks with and without an exhaustively defined set of classes. International Journal of Remote Sensing, 25 (15), 3091–3104. [ Google Scholar ] Gunda, N. S. K., Gautam, S. H., & Mitra, S. K. (2019). Artificial intelligence based mobile application for water quality monitoring. Journal of the Electrochemical Society, 166 (9), Article B3031. [ Google Scholar ] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770–778). https://openaccess.thecvf.com/content_cvpr_2016/html/He_Deep_Residual_Learning_CVPR_2016_paper.html Hernandez-Guedes, A., Santana-Perez, I., Arteaga-Marrero, N., Fabelo, H., Callico, G. M., & Ruiz-Alzola, J. (2022). Performance evaluation of deep learning models for image classification over small datasets: Diabetic foot case study. IEEE Access, 10 , 124373–124386. [ Google Scholar ] Ighalo, J. O., Adeniyi, A. G., & Marques, G. (2021). Artificial intelligence for surface water quality monitoring and assessment: A systematic literature analysis. Modeling Earth Systems and Environment, 7 (2), 669–681. [ Google Scholar ] Jan, M., Hassan, S., & Ara, A. (2022). Freshwater biodiversity: indicators of eutrophication and pollution. Biodiversity of Freshwater Ecosystems , 263–284. 10.1201/9781003277125 Kensington, C. G., Richards, K., Murray, D. L., & Peake, B. M. (2004). Distribution of nitrogen species in groundwater in the Taieri Plain aquifer, New Zealand. Journal of Hydrology (New Zealand), 43 (1), 39–57. http://www.jstor.org/stable/43944913 Khalid, H. (2024). Modern techniques in detecting, identifying and classifying fruits according to the developed machine learning algorithm. Journal of Applied Research and Technology, 22 (2), 219–229. [ Google Scholar ] LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521 (7553), 436–444. [ DOI ] [ PubMed ] [ Google Scholar ] Levin, R. B., Epstein, P. R., Ford, T. E., Harrington, W., Olson, E., & Reichard, E. G. (2002). US drinking water challenges in the twenty-first century. Environmental Health Perspectives, 110 (suppl 1), 43–52. [ Google Scholar ] Lorenzo, P., Luca, P., & Giorgio, G. (2020). Convolutional neural networks for relevance feedback in content based image retrieval. Multimedia Tools and Applications, 79 (37–38), 26995–27021. [ Google Scholar ] Majumdar, D. (2003). The blue baby syndrome: Nitrate poisoning in humans. Resonance, 8 (10), 20–30. [ Google Scholar ] Malathi, V., Manikandan, A., & Krishnan, K. (2023). Optimzied resnet model of convolutional neural network for under sea water object detection and classification. Multimedia Tools and Applications, 82 (24), 37551–37571. [ Google Scholar ] Mishra, R. K. (2023). Fresh water availability and its global challenge. British Journal of Multidisciplinary and Advanced Studies, 4 (3), 1–78. [ Google Scholar ] MohammadiNasab, P. (2023). Pneumonia detection using deep convolutional neural networks . https://www.researchgate.net/profile/Poorya-Mohammadinasab/publication/373653509_Pneumonia_Detection_Using_Deep_Convolutional_Neural_Networks/links/64f647fd4c70687b8ed102be/Pneumonia-Detection-Using-Deep-Convolutional-Neural-Networks.pdf Muhammad Roman, A. Z., Nafchi, A. M., & Sher, M. (2026). Development of an AI-enabled vision-based on-site rapid detection of nitrates and nitrites from drainage samples. Water, Air, & Soil Pollution. 10.1007/s11270-026-09143-7 [ Google Scholar ] Mushtaq, N., Singh, D. V., Bhat, R. A., Dervash, M. A., & Hameed, O. B. (2020). Freshwater contamination: Sources and hazards to aquatic biota. In Fresh water pollution dynamics and remediation (pp. 27–50). Springer Singapore. 10.1007/978-981-13-8277-2_3 Narayana, B., & Sunil, K. (2009). A spectrophotometric method for the determination of nitrite and nitrate. European Journal of Analytic Combinatorics, 4 (2), 204–214. [ Google Scholar ] Picetti, R., Deeney, M., Pastorino, S., Miller, M. R., Shah, A., Leon, D. A., Dangour, A. D., & Green, R. (2022). Nitrate and nitrite contamination in drinking water and cancer risk: A systematic review with meta-analysis. Environmental Research, 210 , Article 112988. [ DOI ] [ PubMed ] [ Google Scholar ] Pu, F., Ding, C., Chao, Z., Yu, Y., & Xu, X. (2019). Water-quality classification of inland lakes using Landsat8 images by convolutional neural networks. Remote Sensing (Basel), 11 (14), Article 1674. [ Google Scholar ] Revsbech, N. P., Nielsen, M., & Fapyane, D. (2020). Ion selective amperometric biosensors for environmental analysis of nitrate, nitrite and sulfate. Sensors, 20 (15), Article 4326. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Saxena, V. (2025). Water quality, air pollution, and climate change: Investigating the environmental impacts of industrialization and urbanization. Water, Air, & Soil Pollution, 236 (2), 1–40. [ Google Scholar ] Shazia, A., Xuan, T. Z., Chuah, J. H., Usman, J., Qian, P., & Lai, K. W. (2021). A comparative study of multiple neural network for detection of COVID-19 on chest X-ray. EURASIP Journal on Advances in Signal Processing, 2021 , 1–16. [ Google Scholar ] Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 . Sumaya, A. I., Forhad, S., Rafi, M. A., Rahman, H., Bhuyan, M. H., & Tareq, Q. (2024). Comparative analysis of AlexNet, GoogLeNet, VGG19, ResNet50, and ResNet101 for improved plant disease detection through convolutional neural networks. In 2024 2nd International Conference on Artificial Intelligence, Blockchain, and Internet of Things (AIBThings) (pp. 1–6). 10.1109/AIBThings63359.2024.1086340 Topping, M., & Kolok, A. (2021). Assessing the accuracy of nitrate concentration data for water quality monitoring using visual and cell phone quantification methods. Citizen Science: Theory and Practice , 6 (1). 10.5334/cstp.346 Wu, W., Huo, L., Yang, G., Liu, X., & Li, H. (2025). Research into the application of ResNet in soil: A review. Agriculture (Basel), 15 (6), Article 661. [ Google Scholar ] Yan, T., Shen, S.-L., & Zhou, A. (2022). Indices and models of surface water quality assessment: Review and perspectives. Environmental Pollution, 308 , Article 119611. [ DOI ] [ PubMed ] [ Google Scholar ] Yang, H., Ni, J., Gao, J., Han, Z., & Luan, T. (2021). A novel method for peanut variety identification and classification by improved VGG16. Scientific Reports, 11 (1), Article 15756. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Yang, L., Driscol, J., Sarigai, S., Wu, Q., Lippitt, C. D., & Morgan, M. (2022). Towards synoptic water monitoring systems: A review of AI methods for automating water body detection and water quality monitoring using remote sensing. Sensors, 22 (6), Article 2416. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Yuesheng, F., Jian, S., Fuxiang, X., Yang, B., Xiang, Z., Peng, G., Zhengtao, W., & Shengqiao, X. (2021). Circular fruit and vegetable classification based on optimized GoogLeNet. IEEE Access, 9 , 113599–113611. [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement Data will be provided on request. Articles from Environmental Monitoring and Assessment are provided here courtesy of Springer ACTIONS View on publisher site PDF (4.4 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