We curated a balanced set of 21,000 high-resolution leaf images covering four major crops, namely corn, potato, rice, and tomato. Healthy and diseased samples are both included, yielding 14 classes in total. The diseases comprise leaf spot, common rust, and northern leaf blight in corn; early and late blight in potato; bacterial blight, brown spot, and leaf smut in rice; and bacterial spot and spider mite damage in tomato [1–3].

Because healthy and diseased samples are equally represented for every crop, this design reduces bias and enables reliable macro-level analysis. Representative examples are shown in Figure 2 and Table 1. Fixed train, validation, and test splits were used to ensure reproducibility and prevent data leakage. Cross-validation was not applied to maintain consistent evaluation across all models.

All experiments were conducted using Python 3.10 with TensorFlow 2.15 and Keras 2.15. The models were trained on an NVIDIA GeForce RTX 3060 GPU (12 GB VRAM) using CUDA 12.2 and cuDNN 8.9 acceleration. Image preprocessing and augmentation were implemented using OpenCV 4.8 and TensorFlow image processing utilities. Data augmentation was applied uniformly across all models, including random rotations of up to ± 15°, horizontal and vertical flips, and brightness adjustments of ± 10% to improve model generalization. All models were trained using the same data splits, preprocessing procedures, and evaluation protocols to ensure a fair and reproducible comparison. Model training and evaluation were performed on a workstation equipped with an Intel Core i7 processor and 32 GB RAM.

We designed a foliar-specific CNN to capture multi-scale lesion cues under diverse conditions. The proposed backbone consists of five convolutional blocks, where the input size is 224 × 224 × 3. The number of filters in each block is as follows: Block 1: 32 filters, Block 2: 64 filters, Block 3: 128 filters, Block 4: 256 filters, and Block 5: 512 filters. The flattened feature maps are then connected to a dense layer followed by a 14-way Softmax output. The total number of trainable parameters in the model is approximately 2.3 million. Regularization techniques include dropout (0.25), label smoothing, and weight decay, while optimization uses categorical cross-entropy with a staged learning-rate schedule and early stopping. With input (W₁, H₁) and convolution parameters (F, P, S), where F is the filter size, P is the padding, and S is the stride, the output width W₂ and height H₂ are given by:

(1) W 2   =   W 1   –   F   +   2   ×   P S   +   1

(2) H 2   =   H 1   –   F   +   2   ×   P S   +   1

and pooling with window size F_s and stride S_s reduces the spatial size as:

(3) W 2   =   W 1   –   F s S s   +   1

Feature maps are flattened and passed to a dense layer, then a 14-way Softmax:

(4) p i   =   e z i ∑ j   =   1 14 e z j

for class probabilities, where z_i is the logit of class i, z_j is the logit of class j (j = 1, …, 14), and p_i is the predicted probability of class i. Regularization includes dropout (e.g., 0.25), label smoothing, and weight decay; optimization uses categorical cross-entropy with a staged learning-rate schedule and early stopping. The architecture’s progressive depth and controlled down sampling stabilize small-lesion detection while preserving context, which—within our unified pipeline—drives its superior precision/recall across all crops-disease categories.

ResNet-50 employs residual learning to stabilize very deep optimization. Instead of directly approximating the target mapping H(x), each residual block learns a residual function F(x) and adds it to the input via an identity shortcut,

(5) F x   =   H ( x )   –   x

which yields through the skip connection:

(6) H x   =   F x   +   x

These skip connections preserve gradient flow and mitigate vanishing/exploding gradients in 50-layer stacks. Our configuration ingests RGB inputs of 224 × 224 × 3, retains the standard stem and fine-tunes all convolutional stages from ImageNet initialization.

VGG-16 stacks small 3 × 3 convolutions to build depth efficiently, interleaved with max-pooling to down sample. Given input size (W_in, H_in), a convolution with kernel F, padding P, and stride S yields the output size (W_out, H_out):

(7) W o u t   =   W i n   –   F   +   2 P S   +   1

(8) H o u t   =   H i n   –   F   +   2 P S   +   1

followed by max-pooling that halves spatial dimensions. We fine-tune an ImageNet-pretrained VGG-16 on 224 × 224 × 3 leaf images, replacing the way head with a 14-class Softmax layer and retaining dropout in the fully connected blocks. To curb overfitting on leaf imagery, we use L2 regularization, moderate data augmentation, class-balanced sampling, and early stopping. This design captures fine textures (speckles, rust pustules) in early blocks and progressively aggregates larger lesion structures in deeper layers, yielding a strong, reproducible baseline under our unified training protocol.

SVM separates classes by maximizing the margin in a (possibly) high-dimensional feature space via kernels. Specifically, the features consist of raw pixel values from resized leaf images of 224 × 224. These pixel vectors are then reduced using principal component analysis (PCA) to the top 20 components and standardized using StandardScaler. This PCA-reduced, standardized representation is used as input for SVM, k-NN, and Random Forest classifiers to ensure reproducible and consistent evaluation. We tune C, kernel [linear vs. radial basis function (RBF)], and γ (for RBF) using GridSearchCV under a fixed training/validation split. This controls bias-variance trade-offs and improves generalization to field variability. Model selection is driven by precision, recall, F1-score, and inspection of the confusion matrix.

Decision function:

(9) f x   =   s i g n ∑ i   =   1 n α i y i   K x i ,   x   +   b

Here x is the input feature vector, x_i (i = 1, …, n) are the n training samples, K( , ) is the kernel, α_i are Lagrange multipliers, y_i are labels, and b is the bias. Linear kernels favor speed and interpretability; RBF captures nonlinear lesion textures and backgrounds. We persist the fitted SVM together with PCA and scaling parameters for reproducible, real-time inference [17].

Random Forest ensembles many decorrelated decision trees, each trained on a bootstrap sample with random feature subsetting, yielding robust performance on heterogeneous leaf imagery. The input features for each decision tree consist of the same PCA-reduced and standardized raw pixel vectors from resized leaf images (224 × 224) used for SVM and k-NN. This ensures consistent representation across traditional classifiers and improves interpretability of feature importance scores [17].

Ensemble vote:

(10) y ^ x   =   m o d e   { f 1 x ,     f 2 x ,   . . . . . ,     f t x }

Here f_t(x) is the class predicted by the t-th tree (t = 1, …, T), and T is the number of trees. Majority voting reduces variance relative to single trees, improving stability under variable lighting, leaf pose, and lesion size. Trained models—and their preprocessing artifacts—are serialized for consistent deployment [17].

k-NN classifies a test sample by the majority class among its k closest training neighbors in feature space. Specifically, the features consist of raw pixel values from resized leaf images (224 × 224), which are then reduced to the top 20 principal components using PCA and standardized with StandardScaler. This representation is used to compute distances for k-NN classification, ensuring reproducibility and consistency with other traditional classifiers [17].

Distance (Euclidean):

(11) d u ,   v   =   ∑ j   =   1 m u j   -   v j 2

Here u and v are m-dimensional feature vectors, and u_j and v_j are their j-th components. The predicted label is the majority vote among the k nearest v to u. While inference scales with dataset size, k-NN provides a transparent baseline and strong performance when features are well-separated. We export the trained pipeline (PCA, scaler, classifier) for reliable, real-time classification [17].

Early stopping based on the validation set is used to control the training of every model. The final results on the held-out test set are reported with the standard metrics given below:

(12) A c c u r a c y   =   T r u e   P o s i t i v e   +   T r u e   N e g a t i v e T r u e   P o s i t i v e   +   T r u e   N e g a t i v e   +   F a l s e   P o s i t i v e   +   F a l s e   N e g a t i v e

(13) P r e c i s i o n   =   T r u e   P o s i t i v e   T r u e   P o s i t i v e   +   F a l s e   P o s i t i v e

(14) R e c a l l   =   T r u e   P o s i t i v e   T r u e   P o s i t i v e   +   F a l s e   N e g a t i v e

(15) F 1   =   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

Because the same training criteria, data splits, and preprocessing are applied to every model, the contribution of each architecture can be examined in isolation. The results indicate that the custom CNN clearly surpasses every baseline, followed closely by ResNet-50 and VGG-16, whereas the traditional methods generalize less effectively by comparison [17].

As shown in Table 2, the proposed custom CNN delivered the best overall results, with 99.29% accuracy over all 14 classes, confirming that it recognizes leaf diseases effectively across multiple crops. We attribute this advantage to an architecture tailored to leaf images, which extracts multi-scale texture cues and fine morphological details of small lesions while distinguishing them from natural leaf structures. The model achieved perfect precision and recall for healthy rice leaves, while the F1-scores for tomato bacterial spot and potato late blight exceeded 0.98. Difficult categories with overlapping lesion characteristics, such as corn northern leaf blight and common rust, still reached approximately 97.04% accuracy, further confirming the robustness of the proposed approach. A learning-rate schedule, calibrated regularization, and balanced sampling kept training stable, while preprocessing and augmentation improved generalization to field-like conditions. Data augmentation included random rotations of up to ± 15°, left–right and up–down flips, and ± 10% changes in brightness, applied uniformly across all models. These design decisions account for the consistent advantage of the custom CNN over both the transfer-learning and traditional baselines, with dependable precision and recall for all crops and diseases. While the custom CNN performed exceptionally on the curated test set, these results should be interpreted cautiously. The dataset is balanced and curated, and the model may have encountered visually similar examples during training. Misclassifications were rare, mainly occurring between visually similar classes such as potato late blight versus early-stage healthy leaves and corn northern leaf blight versus common rust, highlighting the challenges of overlapping lesion morphology and variable backgrounds. Comparative evaluation showed that ResNet-50 and VGG-16 achieved 96.75% and 95.63% overall accuracy, respectively, as shown in Table 2, with minor sensitivity issues on low-contrast lesions. The traditional classifiers (k-NN, SVM, and Random Forest) were consistently weaker than the CNN, with overall accuracies between 89.71% and 90.87%, as shown in Table 3. These results underscore the superior feature extraction and robustness of the custom CNN under the given experimental conditions.

ResNet-50 demonstrated high accuracy of 96.75% and high accuracy on tomato bacterial spot with F1 score of 0.9761, precision of 0.9707, and potato early blight with an F1 score of 0.9834 as presented in Table 2. It had the lowest class of rice leaf smut in rice, which had F1 score of 0.9795, precision 0.9850 and its error analysis revealed it confused corn northern leaf blight where low-contrast lesions diminished sensitivity as we observe in Table 2 and Figure 4. These trends are in line with field imaging where mottled backgrounds and initial stages of symptoms squeeze lesion-leaf contrast. Nonetheless, these residual links provided effective convergence and stabilized optimization to produce robust cross-crop adaptability and competitive behavior among foliar morphologies that were different.

VGG-16 was closely behind with an accuracy of 95.63%, which is performing well on corn common rust with F1 score of 0.9781, precision of 0.9640 and rice brown spot with F1 score of 0.9836, but poorer on potato late blight with F1 of 0.8827, with recall of 0.8778, as illustrated in Table 2. The fine texture cues in the architecture presented by 3 × 3 uniform stacks were consistent and effective, but in cases of irregular or diffuse lesion boundaries. This is notwithstanding the limitation, which despite its weaknesses, allowed VGG-16 to have a balance in precision-recall profiles across most categories, which has provided a stable and reproducible baseline. Practically, they can be interpreted as improvements of the combination of stronger augmentation and class-aware sampling with VGG-16 in terms of shape variability, and the usage of ResNet-50 when sensitivity in heterogeneous imaging conditions is prioritized as overviewed in Figure 4.

k-NN attained an overall accuracy of 89.71% with an F1 score of 0.9944 on corn common rust and 0.8674 on tomato spider mite infestations where lesions are small, sparse and irregular, as evidenced in Table 3. Voting based on distance had difficulty separating classes with fine-grained, low-contrast patterns.

SVM was also found to have 90.87% accuracy with F1 of 0.9474 with rice bacterial blight separation, but the recall decreased in potato late blight 0.8370 recall as demonstrated in Table 3. The sensitivity of the margin-based classifier to heterogeneous lesion morphology and variation of backgrounds was also found to be a weak point in issue of sensitivity in field use.

Random Forest also achieved the accuracy of 90.77%, and it predicted the corn Cercospora leaf spot with an F1 of 0.8555 and tomato bacterial spot with F1 of 0.8934, as shown in Table 3. Figure 5 confusion matrices have more overlaps and misclassifications, reflecting the weaknesses of handcrafted features in the presence of high variability and the presence of minor inter-class differences.

The proposed custom CNN demonstrated reliable classification performance on representative test samples, correctly identifying the actual and predicted classes of input leaf images, as illustrated in Figure 6. Across more than one hundred independent test samples, only one misclassification was observed, where a healthy leaf was incorrectly predicted as a diseased class. Specifically, a Tomato Healthy sample was misclassified as Tomato Bacterial Spot, highlighting the difficulty of distinguishing subtle early symptoms from naturally healthy leaf characteristics. Overall, the custom CNN achieved a strong balance between accuracy, robustness, and efficiency for multi-crop leaf disease recognition.

The interface will present every prediction with two clinician-facing components. Reasons, a summarization of the visual evidence and diagnostic indications such as lesion color, shape, distribution, and context that caused the model to make that decision; and Solutions, which will transform that diagnosis into management steps, which are sanitation, cultural practices, resistant varieties, and where appropriate, chemistry with mode-of-action rotation. The system assists in triage based on thresholds and coherent and defensible decisions in the field by combining calibrated probabilities with open and transparent reasons and guidance.

The trained model was deployed in a web application to demonstrate real-time inference. While a screenshot illustrates the interface and workflow, quantitative evaluation of live performance was not conducted; the system provides a framework for rapid leaf disease assessment in future practical use. The Multi-crop Leaf Disease Recognition system has an end-to-end workflow represented in Figure 7. The landing page (Figure 7a) provides brief instructions and a sample image, which directs the users to a one-session route, starting with uploading to advise. The ingestion view (Figure 7b) allows drag-and-drop/file browsing of JPEG/PNG images, which makes phone/laptop submission straightforward. The results view (Figure 7c) displays the best result of a calibrated confidence of the top class, the original image to be visually checked and Reasons and Solutions could be read by humans. The file can be cleared, and analyses can be re-run immediately. Standardized preprocessing and CNN inference provide consistency of inputs, constant latency, and predictable outputs under the hood. Developed to be used in checkpoints, the system is designed to be extended into clinics and used on the farm, so the model is both accurate and usable without handing over a lot of misclassified data to the decision-makers.