The dataset was sourced from the SAR-CLD-2024 dataset, which contains high-quality images of cotton disease. Dataset of 2137 images from the NCRI (National Cotton Research Institute), Gazipur. The images are taken by a smartphone (Redmi Note11s). This robust dataset covered 7 different classes, including both biotic and abiotic stresses.

The leaves from all 7 classes are illustrated in Fig. 2, and the names of the cotton diseases and their corresponding images are shown in Table 2.

Fig. 2

The individual image of the seven classes. (A) Healthy Leaf, (B) Bacterial Blight, (C) Curl Virus, (D) Leaf Variegation, (E) Jassids by Leaf Hopper, (F) Red Leaf and (G) Herbicide Growth Damage.

Table 2

Number of images in individual class.

To apply the YOLO classification model to the obtained dataset, the data must be organized into three folders: “train”, “val”, and “test”. Each folder contains seven subfolders, each named after one of the seven classes, with the corresponding images for that class. Figure 3 illustrates the format used by YOLO to classify the dataset.

Fig. 3

Format of dataset for YOLO classification.

Several preprocessing steps are applied to ensure consistency in the model’s learning efficiency. All images were resized to 640 × 640 pixels, matching the input size of the YOLOv8architecture.The dataset was split using Python, with the data randomly divided into training (69%), validation (12%), and testing (19%) sets. A significant number of images were allocated for testing to assess the accuracy of the trained model. Table 3 outlines the distribution of the dataset into training, validation, and testing subsets for each individual class.

Table 3

Distribution of images in train, validation, and test.

This technique is applied to divide the dataset into ‘k’ parts, named ‘folds’, to carry out a more accurate method of model performance. Each fold provides data for both training and validation. The k-fold cross-validation applied to object classification scenarios ensures robustness to the DL model, making it perfectly generalizable for various data splits. Cross-validation is particularly important in agriculture, as environmental conditions may vary, causing the appearance of leaves and disease symptoms to differ from those seen in the training set^33,34 (Sohail et al., 2023; Samuel et al., 2024). K-fold cross-validation, combined with DL architectures such as CNN and ResNet-152V2, has been shown to improve the predictive capabilities of the model for classifying and diagnosing cotton plant diseases, thus enhancing its effectiveness in real-world applications (Jai Vignesh et al., 2023)³⁵. Training dataset frequently suffers from overfitting,i.e., reduced performance on new, unknown images. K-fold cross-validation addresses this problem by evaluating the performance of the model across different data partitions. It ensures that the model does not simply memorize the training data but instead learns to generalize (Gayatri et al., 202)³⁶. For further enhancement of the model’s robustness, a largely diversified dataset for cross-validation is used³⁴ (Samuel et al., 2024; Kumar et al., 2024)^34,37. To further strengthen the model, we employed the k-fold technique, creating ten distinct training, validation, and test folds. Each fold was randomly split to ensure variability in the dataset, with the random splitting and fold formation implemented using Python programming. An example of a k-fold process is shown in Fig. 4. The dataset is split into three categories, and the same process is done for ten different Folds. For each of the ten folds, the dataset was divided into three categories, with each fold containing a unique set of images.

Fig. 4

K-fold splitting of the dataset.

The output of an image classifier consists of a single class label and a confidence score. Image classification is particularly useful when the goal is to identify the class to which an image belongs without needing to pinpoint the exact location or shape of the objects within it. YOLOv8 models, specifically the yolov8m-cls.pt variant (Fig. 5) is designed for efficient image classification. The model assigns a class label and a confidence score to an entire image. This approach is especially valuable in applications where knowing the class of an image is sufficient, rather than requiring detailed information about the location or shape of objects it contains.

Fig. 5

Shows the detailed architecture of the YOLOv8 classification model.

The YOLOv8m-cls model contains 141 layers, 15,781,303 parameters, 15,781,303 gradients, and 41.9 GFLOPs. Out of these, we used 103 layers, 15,771,623 parameters, 0 gradients, and 41.6 GFLOPs. NVIDIA GeForce RTX 3050 Ti Laptop GPU, 4096MiB and Intel i7 12th gen processor were used to perform the desired experiment. Initial hyperparameters {Ir0 = 0.01, momentum = 0.937, Irf = 0.01, wgt_decay = 0.0005, warmup_epochs = 0.0005, warmup_decay = 3.0, warmup_momentum = 0.8, warmup_bias_Ir = 0.1, box = 7.5, cls = 0.5, dfl = 1.5, pose = 12.0, kobj = 1.0, label_smoothing = 1.0, label_smoothing = 0.0, and nbs = 64} have been used.

Figure 6 illustrates the augmentation strategies of the YOLOv8 model. Default parameters {hsv_h = 0.015, hsv_s = 0.7, hsv_v = 0.4, degrees = 0.0, translate = 0.1, scale = 0.5, shear = 0.0, perspective = 0.0, flipud = 0.0, fliplr = 0.5, bgr = 0.0, mosaic = 1.0, mixup = 0.0, copy_paste = 0.0, auto_augment: randaugment, erasing = 0.4 and crop_fraction = 1.0} has been used.

These augmentation techniquesaddress the class imbalance present in the SAR-CLD-2024 dataset. It increases the representation of minority classes and helps the model to learn balanced features, which improves the generalization and reduces class-wise prediction bias.Additionally, the use of 10-fold cross-validation ensured, all classes were fairly represented across training and validation splits.

Fig. 6

Augmentations of YOLO model shows the different types of augmentations used internally by the YOLOv8m classify model to classify the leaf for example the defalt settings are {hsv_h = 0.015, hsv_s = 0.7, hsv_v = 0.4, degrees = 0.0, translate = 0.1, scale = 0.5, shear = 0.0, perspective = 0.0, flipud = 0.0, fliplr = 0.5, bgr = 0.0, mosaic = 1.0, mixup = 0.0, copy_paste = 0.0, auto_augment: randaugment, erasing = 0.4, and crop_fraction = 1.0}.

Figure 7 illustrates the plots depicting losses (both training and validation) and Top_1 and Top_5 accuracy for 100 epochs. The losses decreased and stabilized at 0.1 and 1.2 for training and validation, respectively. The value of these losses demonstrates that effective learning with minimum overfitting is achieved. The Top_1 accuracy exhibits a rapid increase from approximately 80% to around 99%, demonstrating the strong ability of the model to predict the correct class on the first attempt. The Top_5 accuracy remains consistently at 1.0, signifying that the model consistently includes the correct label within its Top_5 predictions. Table 4 illustrates trial metrics, and Table 5 illustrates best trials.

Fig. 7

Best trial validation results show the four graphs, including two metrics and two loss graphs. All four graphs show excellent results; both losses are equivalent to zero. Top_1 and Top_5 accuracy of 99.60% and 100% is achieved respectively.

The 100% Top-5 accuracyachieved by the model is expected in this case due to the limited number of classes (8) and strong performance of the trained model. Since Top-5 accuracy only checks whether the correct label appears in the top five predictions, such results are common when the model learns well-separated features. However, Top-1 accuracy remains the primary indicator of model effectiveness, as it reflects the model’s ability to correctly predict the disease in a single attempt.

A consistently high performance with 98.41% accuracy, 98.39% precision, 98.53% recall, and 98.42% of F1_Score is achieved throughout the 10 trials as illustrated in Table 4. The second trial yielded the best results, achieving the highest accuracy at 99.60%, while the other trials also had strong performance. This suggests that the model is robust (Fig. 10), with minor variations in the results likely due to differences in conditions (Figs. 11, 12, 13, 14).

Despite achieving high accuracy values (Top-1: 99.60%, Top-5: 100%), the proposed model does not suffer from overfitting. This conclusion is supported by multiple observations drawn from model behavior and data characteristics. Firstly, the model was evaluated using 10-fold cross-validation, ensuring that each subset of data is used for both training and validation. The performance remained consistentacross all folds, which reflects the robustness and generalizability of the model.Secondly, the confusion matrix generated during validation shows minimal misclassifications, which confirms that the model maintains its classificationabilities on unseen data.Also, the SAR-CLD-2024 dataset, without any augmentations, containing real-worldunique images, is used to train the model. No synthetic data or repetition was used during training, which guarantees that the model has learnt diverse and realistic field conditions.

Principal Component Analysis (PCA) was performed on the deep feature vectors extracted from the final layer of the YOLOv8 classifier. As shown in Fig. 8, the embeddings from different classes have formed distinct and well-separated clustersin the plot. This evidence confirms that the model has effectively learnt discriminative features and is not merely memorizing the training data.

Fig. 8

Principal Component Analysis (PCA) on the deep feature vectors extracted from the final layer of the YOLOv8 classifier.

To further evaluate the effectiveness of the proposed YOLOv8 model, a comparative analysis with YOLOv11, which is a recently released version of the YOLO architecture, is conducted. Both models were trained and validated on the same dataset using identical parameters, including batch size, epochs, and input resolution.

As shown in Fig. 9, YOLOv8 consistently outperformed YOLOv11 in key performance metrics, includingtrain_loss, val_loss, top1 accuracy and top5 accuracy. This suggests that although YOLOv11 is a newer version in the YOLO series, it may not yet be fully optimised for image classification tasks, particularly in the context of fine-grained agricultural disease detection.

Fig. 9

Comparative analysis between YOLOv8 and YOLOv11.

Comparison of classification metrics between YOLOv8 and YOLOv11 on cotton leaf disease classes.YOLOv8 demonstrates smoother convergence, supporting its use for the proposed method.Our experiments revealed that YOLOv11 struggled to achieve stable convergence, as shown in Fig. 9, with fluctuating loss curves and lower accuracy.In contrast, YOLOv8 offersa well-balanced architecture and consistent results across multiple datasets, especially in our use case, making it more suitable for deployment in real-world agricultural scenarios.It is also important to note that YOLOv9 and YOLOv10 do not provide support for image classification tasks, which further supports the selection of YOLOv8 for our study.These findings justifythe selection of YOLOv8 in our study over newer yet less stable alternatives like YOLOv11.

Table 4

Metrics of all ten trials.

Table 5

Metrics of best trial.

Table 5 highlights the peak performance of the DL model during its most successful trial in diagnosing cotton diseases. In this best trial, the model achieved a Top_1 accuracy of 99.60%, indicating that it correctly identified the disease as its top prediction nearly every time. The Top_5 accuracy remained at 100%, ensuring that the correct diagnosis was always included within the top five predictions. The recall was 99.55%, demonstrating the exceptional ability of the model to correctly identify nearly all actual disease cases, minimizing the likelihood of missed diagnoses. With a precision of 99.53%, the model demonstrated that nearly all of its positive predictions were accurate, effectively reducing the number of false positives. F1_Score of 99.60% balanced out the precision and recall results, proving the efficiency of the model in cotton disease detection. This best trial underscores the superior performance of the model, which shows its potential as a highly reliable tool for precision agriculture (Figs. 10, 11, 12, 13 and 14).

Fig. 10

Accuracy vs. epochs of each trial: shows the accuracy Vs epochs for each trial, and the dark blue line shows the average of all ten trials.

Fig. 11

Graph of F1-score for each trial: shows the bar chart of the F1-score in each trial. The best trial was observed to be trial 2.

Fig. 12

Graph of precision for each trial: shows the bar graph of the precision of each trial. Trial two performed the best in all the trials.

Fig. 13

Graph of Recall for each trial: shows the bar chart of the recall of each trial. Trial two showed excellent results of 99.55% in all ten trials.

Fig. 14

Graph of Accuracy for each trial: shows the bar chart of the accuracy of each trial. The best trial was trial two, which showed an outstanding accuracy of 99.60%.

Table 6

Shows the average of all five metrics of ten trials. The average accuracy of ten trials is noted as 98.41%.

Table 6 reveals the strength and high accuracy of the proposed model in diagnosing cotton diseases. The Top_1 accuracy of the model was equal to 98.41%, meaning that in almost all cases, the disease was predicted correctly as the top prediction. The Top_5 accuracy grew to 100%, ensuring the correct disease was always present among the first five predictions and emphasizing the reliability of the model. The model performed pretty well on the test set: 98.53% Recall, meaning it had the ability to effectively identify almost all cases of disease, which minimizes missed diagnoses; 98.39% precision, meaning most positive predictions by the model are correct, thus avoiding false positives. This makes the F1-score 98.42%, indicating that this model is highly effective and consistent over ten separate trials. These indicators support the ability of the model to accurately and reliably diagnose cotton diseases, being of use in precision agriculture.

Figure 15 is a confusion matrix presenting the validation performance of a classification model on different diseased leaves. True labels have been mapped on the x-axis, prediction labels have been mapped on the y-axis, true classifications are illustrated on the diagonal cells, and the off-diagonal cells illustrate misclassifications. This model performs the job extremely well in classification, where it is able to identify “Curl Virus” in 53 out of 53 samples, “Leaf Redding” in 72 out of 72, and “Herbicide Growth Damage” in 34 out of 34. However, out of the 263 samples, the model made just one confusion between the classes “Healthy Leaf” and “Bacterial Blight”. Figure 14 illustrates the normalised confusion matrix (best validation) (Fig. 16).

Fig. 15

Confusion matrix of best trial validation shows the confusion matrix of the validation dataset done by the best trial. In this confusion matrix, all the images are correctly classified as the labels given to them, and only a single image was not correctly identified as the true value. Out of 263 images, only one image was not predicted correctly; otherwise, all the predictions were correct. The accuracy of the given matrix is 99.60%.

Fig. 16

Shows the normalized confusion matrix of the best trial. All the predictions are clearly done correctly, and only one image is not detected correctly.

Prediction of the best trial shows the prediction of the validation set of best trial with an accuracy of 99.60%, and all 16 leaves in the above images are correctly classified. The output of the proposed approach shows the class of each cotton diseased leaf in the left corner of each image, as shown in Fig. 17.

Fig. 17

Prediction of the best trial: shows the prediction of the validation set of the best trial with an accuracy of 99.60%, and all 16 leaves in the above images are correctly classified. The output of the proposed approach shows the class of each cotton diseased leaf in the left corner of each image.