Potato foliage is vulnerable to a spectrum of pathogens with markedly different epidemiologies and symptomatology. Late blight (Phytophthora infestans), a fast-spreading oomycete disease, remains the most devastating, initiating water-soaked lesions that rapidly coalesce into necrotic brown areas with a characteristic grayish, downy sporulation under humid conditions, and frequently extending to tubers (Jafar et al., 2024). Early blight (Alternaria solani) typically emerges on older leaves as discrete dark lesions that enlarge with concentric “target-spot” rings, progressing to chlorosis and premature defoliation in warm, humid environments (Potato Disease Types, 2025). Viral diseases such as Potato virus Y (PVY) and Potato leafroll virus (PLRV), both primarily aphid-borne, induce mosaic mottling, leaf curling, and canopy yellowing (PVY), or the diagnostic upward rolling and brittle texture of leaves (PLRV), with attendant losses in yield and tuber quality that depend on cultivar and viral strain (Liu and Wang, 2021). Soil and vascular-invading fungi drive wilt syndromes early dying dominated by Verticillium dahliae and late dying associated with Fusarium spp. leading to progressive wilting, chlorosis, and necrosis that culminate in significant productivity declines. Bacterial threats such as bacterial ring rot cause leaf yellowing, wilting, and vascular browning with corky ring formation, and can persist in soils and on equipment, complicating eradication efforts. While multiple leaf diseases impair crop performance, early and late blight are generally the most consequential for field management decisions; representative phenotypes for Early blight, Late blight, and Healthy leaves are shown in Figure 1.

This potato leaf disease detection and categorization literature review summarizes current methodologies. Radwan et al. (2025) developed a weather-driven pipeline for early and late blight using K-means, PCA, copula analysis and multiple classifiers, with binary Greylag Goose Optimization for feature selection. On a 4,000-record meteorological dataset, the best MLP with selected features reached 98.3% accuracy. This tabular risk-forecasting setup complements image-based screening. Chen and Liu (2025) introduced CBSNet with Channel Reconstruction Multi-Scale Convolution and Spatial Triple Attention, plus a Bat–Lion training strategy for robustness. On a self-built potato leaf image set, CBSNet achieved 92.04% accuracy and 91.58% precision, extracting tiny lesions and blurred edges effectively. Dey et al. (2025) proposed a lightweight CNN tailored for real-time classification, reducing depth and parameters to 204,227 while preserving accuracy on high-resolution potato leaf images. The model attained 98.6% test accuracy and class-wise precision of 0.99 (early blight), 0.98 (late blight), 1.00 (healthy), outperforming VGG16, AlexNet, and ResNet50. Sinamenye et al. (2025) fused EfficientNetV2-B3 with a Vision Transformer to couple local convolutional features with global context. Trained on the Potato Leaf Disease Dataset reflecting field variability, the hybrid reached 85.06% accuracy, improving prior results by 11.43 points. Salihu et al. (2025) built a CNN trained with Adam, using scaling, augmentation, and normalization over a curated set of healthy, early blight, and late blight images. The model achieved 96.88% accuracy, with class metrics including precision 0.76, recall 0.93, F1 0.84 for healthy and near-perfect scores for blight classes. Ala’a (2025) extracted generalized Jones polynomial texture features and classified with SVM on Plant Village potato images. The GJP-SVM pipeline preprocessing, feature extraction, dimensionality reduction, classification reached 98.45% accuracy, showing strong performance from hand-crafted descriptors. Zhang et al. (2025) benchmarked VGG16, MobileNetV1, ResNet50, and ViT, then proposed VGG16S with global average pooling, CBAM attention, and Leaky ReLU to shrink parameters to 15 M. After response-surface hyperparameter tuning, VGG16S achieved 97.87% test accuracy and generalized well on public sets. Kaur et al. (2025) presented PotConvNet, a compact CNN trained on two potato image datasets with resizing, normalization, augmentation, and fixed splits. Reported accuracies were 99.76% (Dataset 1) and 97.78% (Dataset 2), validated by F1, precision, recall, Cohen’s kappa, and ROC AUC. Nur et al. (2025) optimized Inception V3 via transfer learning and targeted fine-tuning of terminal layers on a domain-specific potato leaf set. The approach yielded 97.78% accuracy with precision 98%, recall 98%, F1 98%, offering strong performance with practical efficiency. Shah et al. (2025) introduced PLDC-Net, using EfficientNet-B1 as a backbone, fine-tuned with dense layers and an SVM output head; data balancing and augmentation were emphasized. On an unseen test set, the model achieved 98.39% average accuracy, providing a reliable transfer-learning baseline for multi-disease identification. We diagnose and categorize potato plant diseases. Various studies on diagnosing and categorizing potato plant diseases may be found in the literature on potato leaf disease classification and detection (Fuentes et al., 2017). CNN and other deep learning approaches have shown promise for automating the detection and classification process, reducing the need for human expertise-several CNN architectures, transfer learning, feature extraction, and ensemble methods to improve accuracy and robustness (Geetharamani and Pandian, 2019). The study (Ahmed et al., 2025) suggested a deep CNN model to identify outstanding and ailing foliage across crops. They trained their model using the Plant Village dataset, which includes photos of diseased and healthy leaves and the backgrounds of 38 distinct crop kinds. However, they did not zero in on potato crop illnesses, and the data used to prepare the algorithm in the United States and Switzerland missed Pakistan-endemic infections on potato leaves.

Despite having little data, the scientists used deep learning, specifically CNNs, to identify potato illnesses (Lee et al., 2020). A CNN model was created (Awal et al., 2019) to distinguish between healthy potato leaves and those affected by early or late blight. They also used the regionally targeted Plant Village dataset in their research (Khalifa et al., 2021). In this research, we looked at how well deep learning methods and convolutional neural networks, in particular, might do at identifying diseases on potato leaves. The authors trained a CNN network using a collection of photos of diseased potato leaves. The success of the suggested method in illness detection demonstrates the promise of deep learning for this application area. According to Ghosal et al. (2019), the CNN model has the ability to differentiate between various plant classes. This study, Rathod et al. (2020) used deep learning to detect potato leaf blight early. The authors trained a CNN architecture to interpret potato leaf images. The model’s early blight detection highlights deep learning’s potential for potato leaf diseases. The authors examined deep learning and transfer learning for potato disease diagnosis (Liang et al., 2019). Using potato leaf images, the authors updated VGG16, a pre-trained CNN model. Pre-trained CNN models with transfer learning were useful in potato disease detection. A network for identifying and assessing plant diseases was demonstrated in Ferentinos (2018). To distinguish between healthy and diseased plants from photographs of their leaves (Rozaqi et al., 2020; Sanjeev et al., 2020) looked at many deep-learning architectures. These included AlexNet, Overfeat, AlexNetOWTBn, VGG, and GoogLeNet. The authors applied transfer learning to the PlantVillage dataset to identify local agricultural diseases. We developed a CNN model to detect potato plants with early, late, or robust blight. We trained the model using PlantVillage, disease data. FFNNs can distinguish between early, late, and healthy foliage (Barman et al., 2020). They trained and tested their system using PlantVillage. Using a self-built CNN (SBCNN) model, Tiwari et al. (2020) classified potato leaves as early, late, or healthy. The regional PlantVillage dataset improved their model’s accuracy. They did not utilize experimental data to validate their model. Gupta et al. (2019) extracted and classified features using KNN, SVM, a neural network, and a pre-trained VGG19 model using KNN, SVM, and a neural network. PlantVillage has trained the computer to identify early and late blight symptoms on potato foliage. Research demonstrates that CNNs and other forms of deep learning effectively identify and categorize diseases in potato leaves. To further improve the performance of deep learning models, even with minimal training data, practitioners have turned to methods including data augmentation, transfer learning, and fine-tuning pre-trained models. These findings show that deep learning may improve potato disease detection and classification, which is crucial for the crop’s long-term health.

The study, Bappi et al. (2025) provided a novel deep-learning algorithm for potato leaf tissue disease detection using augmentation approaches. Scaling, flipping, and rotating the training dataset enhanced the model’s accuracy. The research (Rahman et al., 2021) examined how different kinds of enhancement may affect deep Convolutional Neural Networks (CNNs) ability to spot illnesses in potato leaves. In this research (Plant Village Dataset, 2024), we applied deep learning models and data augmentation to improve our ability to identify diseases in potato leaves. The authors in the research work developed a deep learning-based method that uses data augmentation techniques to detect potato diseases. They used augmentation methods, including scaling, flipping, and rotating, to upsurge the size of the training dataset. In potato disease identification, training a CNN model on the expanded dataset resulted in high accuracy. Table 1 summarizes augmentation and deep learning studies on potato leaf disease detection and classification. These studies demonstrate the scope of current potato leaf disease identification and categorization efforts. While typical machine learning methods have shown promise, recent research has demonstrated that deep learning, particularly CNNs, may boost accuracy and automation. The proposed study on two-stage PotatoLeafNet CNN architectures will examine their ability to accurately identify and classify potato leaf diseases.

Plant Village Dataset provides high-quality photos of different potato leaves (Mishra and Srivastava, 2019). Healthy, Early and Late Blight were photographed. Because of its availability, researchers have used the Plant Village dataset to simulate potato leaf diseases in the literature. This region-specific dataset includes few training and validation pictures and uneven class distribution. We need a fresh and comprehensive potato leaf dataset to address these research gaps. We curate the new dataset as the Potato Leaf Disease Dataset. Early Blight, containing 1,628 potato images, is the most critical disease affecting potatoes. The subsequent severe risk Late Blight contains 1,424 leaf images. We will examine 1,020 leaf images from the Healthy Next class for model training and testing. The dataset contains a complete 4,072 potato leaf images with three classes. The ratio between training, validation, and testing is 80:10:10. Figure 3 displays the potato leaf images from each of the three categories. Figure 3 presents the distribution of images across three classes of potato leaves: Early Blight, Late Blight, and Healthy. The Early Blight class has the largest number of images, just under 1,800, indicating a higher prevalence or focus on this category within the dataset. Late Blight follows closely, with a count near 1,600 images. The Healthy class has the fewest images, slightly above 1,400, suggesting a lesser representation in the dataset. This visual distribution highlights an imbalanced dataset which may be used for training a machine learning model to classify the health status of potato leaves.

Pre-processing was applied to all images to enhance lesion visibility, suppress background clutter, and standardize inputs prior to learning. Specifically, we performed contrast normalization to mitigate illumination variability, foreground–background separation to isolate the lamina, and spatial normalization to a common resolution. This stage improves the signal-to-noise ratio presented to the network and, in turn, the reliability of feature extraction for downstream classification (Hernandez-Valencia et al., 2020). To reduce storage and I/O overhead without compromising diagnostically salient content, we employed lossless and hybrid compression. Lossless codecs Huffman coding and run-length encoding (RLE) preserve the exact pixel values while exploiting redundancy to shrink file size (Yao et al., 2020). In the hybrid scheme, regions containing disease cues (lesion edges, texture) are preserved losslessly, whereas visually noncritical background is compressed lossily, striking a balance between fidelity and efficiency for large-scale training and deployment (TensorFlow Sequential Data Augmentation, 2025). (Compression is decoupled from resizing, it reduces bytes on disk/transfer, not spatial resolution.) leaf images captured in RGB are converted to grayscale (Gurucharan, 2020). Edge of Caution to recognize the edges in a leaf image and alleviate the irritation, unambiguous evidence is utilized (Powers, 2020). The external designs in leaf images are equal in how they are perceived from the edge. When the upper shape is taken as (p, q), the breadth and the level are (r, s), and these four centers do not settle the bobbing (Li et al., 2022). Each member of the upright hopping square is still a work in progress. The return on investment region is removed using the primary RGB leaf image’s coordinates (p + r, q + s). Finally, the dreaded leaf symbol may be put to rest.

The potato leaf disease dataset was divided into training, validation, and testing sets using 80, 10, and 10% split ratios. Sequential image augmentation procedures on the training set reduced overfitting and increased dataset variation. Rescaling, rotating, modifying shear and zoom ranges, flipping horizontally, adjusting brightness, and moving channels were these tactics. CNN model predictions were improved using Adam optimization with forward and backpropagation. Thus, CNN model output accuracy was ensured. The validation and testing sets contained 20% of the training set, which included images of early, robust, and late blight. The PotatoLeafNet model categorized practice pictures and predicted class labels on the training dataset.

The existing literature on deep learning approaches reveals several challenges, including misdiagnosis of potato leaf identification, variations in potato leaves due to different varieties, and environmental factors. Early detection and management of potato diseases are crucial, but the process is time-consuming, and access to agricultural expertise is limited in rural areas (Alzakari et al., 2025). CNNs have shown remarkable progress in image-based recognition, eliminating the need for extensive image pre-processing and enabling automatic feature selection (Weng et al., 2024). However, the availability of large datasets specifically for potato leaf challenges remains a significant obstacle.

Multiple metrics are used to evaluate the success of a network. Using different task metrics helps represent the network’s ability to solve a given problem. The evaluation metrics can use true positive (TP), false positive (FP), true negative (TN), and false negative (FN).

Classification Accuracy: is determined by the ratio of correct prediction to total predictions.

Precision: Precision determines with what precision the network places images in the positive category. Precision is calculated as follows:

Recall: Recall indicates how many positive images the network recorded. The recall is calculated as follows:

F1-Score: F1-Score is a combination of Precision and Recall. The calculation is as follows:

PotatoLeafNet model for potato leaf detection and classification shown in Algorithm 1. ALGORITHM 1

Input: Potato Leaf Disease Dataset Output: Disease Detection and Classification of Potato LeavesStep1: Acquire the Potato images with Late Blight, Early Blight, and Healthy Step2: Loading the data (X_train,y_train), (X_test,y_test)=image.load_data() Step 3: first stage of PotatoLeafNet for sequential image augmentation model for image augmentation with 6 layers. Each layer is performing RandomFlip, RandomRotation(0.2), RandomZoom(0.2), Rescaling(1./255) Step 4: Put the correct labels on the pictures of the potato leaf images. Step 5: Sort photos into categories using the available class labels from the training and testing datasets. Step 6: Initialize the parameters image size, epochs, batch size, and train and test image labels. Step 7: The Second stage of the PotatoLeafNet Model uses 4 blocks containing Conv2D, Max Pool2D, and GloalAveragePooling2D, followed by Desne layers. The total number of layers is 11. Step 8: Evaluate the trained model using a separate testing dataset.    Calculate the test loss and accuracy of the model. Step 9: Check the accuracy of the proposed models, and see how they stack up against the rest of the CNN models out there. Make predictions on new data predictions = model.predict(new_images).

We curated a diverse, high-quality corpus of potato leaf images spanning Healthy, Early Blight, and Late Blight classes. Training uses the PlantVillage Potato subset, a widely used, fully open benchmark for potato leaf disease recognition, to mitigate its limitations and class imbalance, we additionally compiled a complementary Potato Leaf Disease Dataset with 4,072 images 1,628 Early Blight, 1,424 Late Blight, and 1,020 Healthy. To assess real-world generalization beyond PlantVillage, we conduct cross-dataset validation: models trained on PlantVillage are evaluated, without further tuning, on PlantDoc (Singh et al., 2020) (in-situ scenes with variable lighting, occlusion, and background clutter) and on Shabrina et al. (2023) field collection (uncontrolled conditions; seven potato classes remapped to {Healthy, Early Blight, Late Blight} for comparability). We report Accuracy, Macro-F1, per-class Precision/Recall, Matthews Correlation Coefficient (MCC), and Expected Calibration Error (ECE), and provide confusion matrices and Grad-CAM overlays. Finally, a few-shot field-adaptation ablation (10% labeled field images) quantifies domain shift and the benefit of lightweight adaptation.

All images were prepared for CNN training by converting them to RGB float tensors and resizing uniformly to 224 × 224 pixels. Pixel intensities were normalized to [0,1] to stabilize optimization. The dataset comprised 4,072 potato-leaf images across three classes Healthy, Early Blight, and Late Blight. To enhance the robustness and generalizability of PotatoLeafNet, we applied a fixed-order (sequential) image-augmentation pipeline in Keras on the training split only, thereby increasing appearance diversity while preserving label integrity and class balance. The augmentation sequence consisted of rotation (±25°), width shift (±0.10), height shift (±0.10), shear (0.20), random zoom (up to 0.20), horizontal flip, brightness jitter (0.5–1.0), and channel shift (0.05). Applying this policy expanded the training corpus to 6,000 images, balanced as 2,000 per class, which mitigated class imbalance and improved generalization across diverse disease manifestations. Figure 5 illustrates representative pre-processed images at the target 224 × 224 resolution.

Table 3 illustrates PotatoLeafNet configurations. The dataset shows the PotatoLeafNet model, which changes internal parameters to improve performance during training. The model learns to extract important traits and properly characterize Late, early blight, and Healthy over several epochs. Each epoch’s accuracy and loss statistics show model performance. Accuracy is the percentage of correctly predicted instances concerning actual. Figures 6–9 demonstrate the accuracy of PotatoLeafNet architectures’ potato leaf disease detection and classification.

The two-stage CNN model evaluates PotatoLeafNet on potato leaf disease data. Potato leaf disease datasets employ six-layer sequential image enhancement. A freshly developed and fine-tuned CNN model analyzes the training dataset’s accuracy. The PotatoLeafNet model was optimized with sequential picture augmentation. PotatoLeafNet model training requires a huge sample.

Figure 6 summarizes the learning dynamics of PotatoLeafNet trained for 100 epochs with sequential image augmentation on the potato leaf disease dataset. The training accuracy rapidly increases and saturates at 98.92% with a final training loss of 0.0356, while the validation accuracy stabilizes around 97.53% with a closely aligned validation loss curve. The small gap between training and validation accuracies, together with the monotonically decreasing and non-diverging loss trajectories, indicates that the augmented model generalizes well beyond the training set. On the held-out test set, PotatoLeafNet attains 98.52% accuracy, confirming that sequential image augmentation provides effective regularization and supports highly reliable classification of healthy, early blight, and late blight leaves.

Figure 7 shows the behavior of a single-stage PotatoLeafNet trained for 100 epochs without sequential image augmentation. In this setting, the model reaches 96.17% training accuracy with a training loss of 0.2250 and achieves 96.52% validation accuracy, but the validation accuracy and loss curves exhibit noticeably larger oscillations than in Figure 6. Test accuracy is reduced to 96.01%, i.e., 2.51 percentage points below the augmented two-stage PotatoLeafNet, and the substantially higher training loss further reflects less stable optimization. Comparing Figures 6, 7 demonstrates that sequential image augmentation not only increases training/validation/test accuracy by +2.75/+1.01/+2.51 points, respectively, but also yields smoother validation trajectories and lower loss, highlighting its role as an effective regularizer.

Figure 8 depicts PotatoLeafNet trained for only 10 epochs with sequential image augmentation. Even under this short training regime, the model already reaches 88.22% training accuracy with a training loss of 0.3535, while the validation accuracy rises to 86.91% and the validation loss decreases steadily. The corresponding test accuracy of 88.15% confirms that the augmented model generalizes well even before full convergence. These dynamics indicate that augmentation quickly exposes the network to diverse views of each class, enabling the model to acquire discriminative features early in training and to maintain a small and stable train–validation gap.

Figure 9 presents the same 10-epoch training schedule without sequential image augmentation. In this baseline configuration, the model attains 87.82% training accuracy and 0.3410 training loss, with validation and test accuracies of 85.82 and 86.91%, respectively. Compared with Figure 8, both validation and test accuracies are consistently lower and the gap between training and validation curves is slightly larger, suggesting mild overfitting when the network is trained on a less diverse set of images. The corresponding loss curve also shows a less smooth descent, pointing to reduced robustness of the optimization process. Together, Figures 8, 9 illustrate that, even at an early training stage, sequential image augmentation improves generalization and stabilizes the learning dynamics of PotatoLeafNet.

Table 5 benchmarks the proposed PotatoLeafNet against recent potato-disease studies spanning handcrafted descriptors with classical classifiers (Ala’a), transfer-learned CNNs (Nur et al., 2025; Shah et al., 2025), compact bespoke CNNs (Kaur et al., 2025; Salihu et al., 2025), hybrid CNN–Transformer designs (Sinamenye et al., 2025; Zhang et al., 2025), and a non-image tabular risk-forecasting approach (Radwan et al., 2025). Despite heterogeneity in data sources and class definitions, PotatoLeafNet attains 98.52% accuracy on PlantVillage (Healthy/Early/Late), placing it among the top performers while using a compact 11-layer 3 × 3 convolutional stack and a fixed sequential photometric augmentation policy. Notably, several comparators optimize for different modalities (e.g., meteorological risk factors) or field-like imagery; therefore, results are indicative rather than strictly commensurate, and cross-dataset validation remains essential for assessing real-world robustness.

Under the same training–evaluation protocol, the Proposed model achieves 98.52% accuracy, 98.67% precision, 99.67% recall, 99.16% F1-score, and 1.00 AUC. Relative to ResNet-50 + VGG-16 (97.10, 95.00, 94.00, 94.00%, 0.98), this corresponds to absolute gains of +1.42 pp. accuracy, +3.67 pp. precision, +5.67 pp. recall, +5.16 pp F1, and +0.02 AUC Table 6. Against VGG-16 + MobileNetV2 (94.80, 92.00, 91.00, 91.00%, 0.93), the gains are +3.72 pp, +6.67 pp, +8.67 pp, +8.16 pp, and +0.07. Versus MobileNetV2 (93.20, 91.00, 90.00, 90.00%, 0.92), the improvements are +5.32 pp, +7.67 pp, +9.67 pp, +9.16 pp, and +0.08; and versus Inception-V3 (92.50, 90.00, 89.00, 89.00%, 0.91), they are +6.02 pp, +8.67 pp, +10.67 pp, +10.16 pp, and +0.09. The consistent, largest margin in recall indicates the Proposed model substantially reduces false negatives critical for early disease screening while simultaneously delivering the best precision, F1, and AUC among all baselines.

The classification accuracy of the model on correctly labeled images reflects the strength of the proposed CNN framework in distinguishing between diseased and non-diseased leaf samples. A correctly identified instance refers to an image that the model assigns the appropriate label to, whether the leaf is affected or unaffected. This aspect of performance was evaluated using standard accuracy-based metrics. The model consistently delivered accurate predictions across all categories, showcasing its reliability in handling both training and unseen test samples. Its ability to differentiate between various visual patterns linked to disease manifestations underlines its robustness and generalizability. The successful identification of all leaf conditions confirms the framework’s precision and operational reliability in practical settings. Figure 10 provides a visual representation of the model’s performance in identifying each class correctly, further validating its strength in class-wise prediction and its potential for real-world application in automated plant disease assessment systems.

Collectively, these metrics provide a comprehensive view of class-wise prediction performance. Training the PotatoLeafNet model for 100 epochs yielded strong results: precision of 98.00% for Early blight, 99.00% for Late blight, and 99.00% for Healthy leaves; recall of 100.00, 99.00, and 100.00%, respectively; and F1-scores of 99.00, 99.00, and 99.50% for Early blight, Late blight, and Healthy leaves, respectively, corresponding to a macro-averaged F1 of 99.16%. Taken together, these indicators suggest that PotatoLeafNet accurately predicts and classifies potato leaf disease categories. At the same time, Table 5 indicates that Early and Late Blight can be less reliably predicted under challenging conditions such as out-of-distribution inputs, limited representativeness in the training data, noisy or ambiguous images, and potential overfitting whereas Healthy leaves are generally classified more accurately. Figure 11 presents the confusion matrix for a small held-out test subset after 100 training epochs (Early Blight = 4, Late Blight = 6, Healthy = 2). All instances were correctly identified, with no false positives or negatives, corresponding to 99% accuracy and precision/recall of >0.98 for each class on that subset. However, given the limited sample size, these perfect results should be interpreted cautiously and validated on larger, more diverse datasets to confirm generalization (Table 7).

Potato production underpins global food security, yet yields and quality are threatened by diverse foliar diseases whose early diagnosis is complicated by cultivar heterogeneity, variable symptom expression, and environmental noise, making rapid and accurate detection essential. To address this need, we propose PotatoLeafNet, a two-stage convolutional framework for automated identification of potato leaf conditions. In the first stage, a fixed sequential image-augmentation pipeline expands intra-class variability and mitigates overfitting; in the second, an 11-layer CNN with 3 × 3 kernels learns discriminative morphological and textural representations from the augmented images. Evaluated on the enhanced dataset, PotatoLeafNet achieved an overall accuracy of 98.92%, with complementary performance measures confirming its ability to correctly categorize samples. In comparative analyses, the approach outperformed representative state-of-the-art baselines and consistently predicted Late Blight, Early Blight, and Healthy classes with high reliability. By enabling precise differentiation among these categories, PotatoLeafNet facilitates timely intervention and supports evidence-based disease-management strategies in real-world agronomic settings.