2.1 Deep learning in major food crops

Joseph et al. [10] implemented a deep learning-based plant disease classification approach for rice, wheat, and maize crops. Their results showed that data augmentation effectively reduced overfitting, achieving classification accuracies of 97% for rice, 98% for wheat, and nearly 99% for maize. Among the tested models, Xception and MobileNet achieved 95.8% and 94.6% accuracy, respectively, for maize diseases, while wheat disease classification reached 96.3%. Rice disease classification yielded the highest performance, with Xception achieving 97.2% and InceptionV3 achieving 96.2%. Girmaw et al. [11] applied transfer learning for plant disease detection, initially using a dataset of 1,400 images, which was later expanded to 3,500 to enhance model performance and reduce overfitting. EfficientNetB7 emerged as the top-performing model due to its compound scaling and advanced architectural efficiency.

2.2 Lightweight and real-time plant disease detection

Bhagat et al. [12] applied a lightweight deep learning model for real-time detection of diseases in the pigeon pea, which is optimized for resource-constrained devices. Using the first real-field pigeon pea dataset with 12,232 images across five disease classes, Lite-MDC employed MDsConv for multiscale feature extraction while minimizing computational complexity. With only 2.2 million parameters, it achieved 94.14% accuracy, outperforming ResNet-50 and DenseNet121. The model was tested on publicly available datasets such as Plant Village and Cassava. The lightweight model made it very feasible for deployment on drones and smartphones, making it suitable for real-time use. Obsie et al. [13] designed an enhanced model for mummy berry disease detection in wild blueberries, incorporating the Coordinate Attention mechanism into the base model for such challenges as occlusion, complex backgrounds, and limited data from field collections. It reached a better precision and recall of 96.3% on synthesized augmented images compared to its baseline model, Yolov5s, which detects infected plant parts at varied spatial scales. The CA module strengthened the discrimination of features, while the CIoU loss enhanced the bounding box regression and increased the robustness in field environments. Lightweight Yolov5s-CA ensured real-time applicability, worked effectively with combined synthetic and real datasets, reduced labor-intensive data collection, and has excellent potential in wide deployment over resource-constrained devices and broader plant disease management applications.

2.3 Deep learning applications in leaf disease classification

Nikitha and his colleague [14] developed a soybean leaf disease identification system. Using image segmentation and K-means clustering, the CNN model attained the highest accuracy of 96%, while SVM and KNN achieved 76% and 64%, respectively. The study showed CNN’s superior performance due to its ability to extract features and effectively handle the multi-class classification challenge. Chowdhury et al. [15] presented the detection of plant leaf disease with the help of deep learning techniques in agriculture in Bangladesh. The authors used the Plant Village dataset, which consisted of 17,430 images from bell peppers, tomatoes, and potato leaves. The applied custom CNN model performed well in crop disease classification, achieving 85.31% accuracy, outperforming the other models. The authors developed a mobile application for real-time disease detection to aid Bangladeshi farmers.

Nigar et al. [16] used the EfficientNetB0 deep learning model for the classification of 38 plant diseases. The system achieved the highest accuracy, precision, and recall rates of 99.69%, 98.27%, and 98.26%, respectively. LIME explainable AI was used to provide visual insights into the classifier’s decisions and provide more insight and interpretability. EfficientNetB0 outperformed other models such as MobileNetV2, CNN, and ResNet-50. Finally, they developed a mobile application called “PlantCare" for real-time disease detection, making this approach practical for farmers. Madhurya and Jubilson [17] presented a more effective deep learning framework, YR2S, for the detection and classification of plant leaf diseases with an accuracy of 99.69%. Optimized by Enhanced Rat Swarm Optimizer (ERSO), the YR2S framework included YOLOv7 for identifying diseases with the ShuffleNetV2 classifier. It combined PCFAN for feature extraction, which captured color, shape, size, edge, and morphological features, and uses the Red Fox Optimization Algorithm to segment diseased areas accurately. The proposed framework outperformed existing frameworks, including DenseNet121 and hybrid VGG16-ResNet-50, in terms of both accuracy and computational efficiency when tested on the New Plant Diseases Dataset comprising 87,867 images over 38 classes.

Salam et al. [18] applied deep learning methods for formulating diseases in mulberry leaves, especially for healthy leaves, leaf rust, and leaf spots. The authors enhanced feature extraction and reduced overfitting through transfer learning applied to a dataset containing 1091 images, augmented to 6000 images. The modified MobileNetV3Small attained the highest accuracy of 97.0% among all other models. Xiao et al. [19] proposed the SE-VRNet, a lightweight deep residual network agile in attention mechanisms to support feature extraction for leaf disease recognition. The model achieved top-1 classification accuracies of 99.73% and 95.71% on NewData and SelfData datasets, respectively. The authors enhanced model generalization and robustness using a dataset expansion strategy with increased image sample numbers to 87867. SE-VRNet amplification outperformed other classical architectures like ResNet, VGG16, and GoogLeNet due to the more significant yet compact model size appropriate for mobile deployment. Such findings captured the essence of AI for mobile agriculture based on the understanding of deep learning, lightweight and deep learning efficient models. Salam et al. [20] implemented the MobileNetV3 deep learning algorithm to predict the leaf diseases of mulberry with an accuracy of 94.2% and 96.4% for the leaf rust and leaf spot classification, respectively. Transfer learning with a training set of 6,000 images increased feature extraction processes and reduced overfitting.

From the literature review, it can be concluded that deep learning techniques have not been used for Malabar spinach leaf disease classification. Significant works have been done on staple crops and region-specific applications utilizing efficient architectures, interpretable frameworks and smartphone-based automatic disease detection. This highlights a clear research gap and the need for a domain-optimized, interpretable, and lightweight deep learning framework for underrepresented crops.

3.1 Dataset description

The dataset used for this study was the Malabar spinach leaf dataset, consisting of three categories: healthy, infected by Alternaria, and infected by straw mites. This dataset was established by combining Habiganj Agricultural University’s private collection with a public straw mite dataset obtained from Daffodil International University. The raw corpus contained slightly more than 700 unique leaf images; after applying the augmentation pipeline described in Sect 3, the effective training set contained over 2,100 samples. The entire combined set was split once into training, validation, and test sets at a 70–15–15 ratio. Exactly the same partition was then used throughout all subsequent experiments, including self-supervised and transformer-based models. Each split was prepared in three class-specific folders according to the standard ImageFolder arrangement. Model pipelines like SimSiam or SwinV2-Base acted on this very same 70–15–15 split, their variations solely coming from their respective internal pretraining or augmentation, not from any alternative split of the dataset. Sample images of the collected spinach leaf disease dataset are illustrated in Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Sample images of the collected spinach dataset.

https://doi.org/10.1371/journal.pone.0340989.g001

3.2 Data preprocessing and augmentation

Data preprocessing is necessary for enhancing image features and improving model performance by removing unwanted objects and standardizing dataset quality. Fig 2 illustrates sample training images after applying the image augmentation techniques. Images were resized to 224×224 pixels, with a focus on cropping diseased leaf areas. Increased model robustness is achieved by applying different augmentation techniques only on the training images, such as horizontal flipping, rotation, brightness adjustment, contrast enhancement, adding Gaussian noise, and perspective transformations. Enhancement operations are used to increase clarity and decrease noise, which includes sharpening and Gaussian blur. Normalization of pixel values was set to a range between 0 and 1, and color space manipulation, such as grayscale and HSV, was opted for to enhance the discriminative features. It mainly uses Python and ImageJ for data cleaning and reliability; therefore, the class imbalance problem could be dealt with through techniques such as oversampling or undersampling. Fig 3 illustrates the class distribution of the Malabar spinach leaf dataset in percentage, showing that healthy samples constitute the largest portion (39.9%), followed by straw mite (31.9%) and Alternaria (28.1%), indicating a moderately imbalanced dataset. Test-time augmentation is applied only to the test split at inference time, and no augmented view of any test image is ever used during training.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Augmented training images of the employed dataset.

https://doi.org/10.1371/journal.pone.0340989.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Percentage-wise class distributions of the employed dataset.

https://doi.org/10.1371/journal.pone.0340989.g003

3.3 Applied deep learning models

Various state-of-the-art pretrained deep learning models, i.e., MobileNetV2, EfficientNetB0, ResNet-50, and ViT-B/16, were used for feature extraction and additional fine-tuning in this research. EfficientNetB0 [21] employs a compound scaling approach, which effectively balances accuracy and computational cost. ResNet-50 [22], a deep residual network, employs skip connections to facilitate enhanced feature learning. ViT-B/16 [23], a transformer-based model, takes images in patch format and employs self-attention mechanisms. All models were initialized with ImageNet weights, their final classifier layer was replaced with a three-class softmax classifier, and they were fine-tuned with appropriate optimizers and learning rates.

3.3.1 Spinach CNN.

The custom SpinachCNN model, illustrated in Fig 4, comprises an initial convolutional layer of 7×7 with a stride of 2, Batch Normalization, ReLU, and Max Pooling. The network is then followed by three residual blocks, each increasing the feature channels from 64 to 128, 128 to 256, and 256 to 512, respectively. A global pooling layer was equipped with average and max pooling for feature extraction. Two fully connected layers follow this step: batch normalization, ReLU activation, and dropout. Finally, a softmax-activated output layer classifies the input into three categories.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Proposed SpinachCNN architecture.

https://doi.org/10.1371/journal.pone.0340989.g004

The SpinachCNN model was designed and trained using an optimal and reliable experimental setup. For the process of optimization, AdamW optimizer was utilized, where the base learning rate was , and the weight decay value was , and it had a batch size of 32 for 30 epochs. In the present code, no learning rate scheduler and early stopping were applied, which trained the code from start to finish for the total number of epochs. For the process of optimization, Cross-Entropy Loss was employed, which utilized label smoothing of 0.1, and mixed precision training with AMP gradient scaling. In the data pipe, extensive augmentation was performed during the process, which involved randomized converted crops, flipped images, rotated images, jitter, and affine transforms, where the images were converted based on the standard mean and standard deviation of the dataset. In the code for an optimal and reliable experimental setup, the random seed was set to 42 for the entire Python, NumPy, and Torch.

3.3.2 Spinach Res-SENet model.

Spinach Res-SENet model, shown in Fig 5, classifies leaf images efficiently using residual connections and squeeze-and-excitation (SE) blocks [24], which makes it possible to represent the features robustly. This architecture consists of an input layer that accepts three-channel images of 224×224, which are preprocessed before entering this network. The initial stage of this model comprises the Convolution Block, which performs a 2D convolution with a kernel size of 7×7 and 64 output channels. Next, batch normalization and a ReLU activation function are applied for non-linearity. A max-pooling layer downsamples the spatial size to obtain only coarse-level features while reducing the computational cost.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Proposed Res-SENet architecture.

https://doi.org/10.1371/journal.pone.0340989.g005

Three Residual Blocks form the backbone of the proposed Res-SENet architecture, each consisting of a primary path and a skip connection. Residual connections allow the model to learn residual mappings; thus, deeper architectures are possible without performance degradation. Each residual block is supplemented with an SE module, which adaptively recalibrates channel-wise feature responses by exploiting the interdependencies among channels to enable the model to perform more accurate prioritization between the different channels.

Residual Block 1 contains two 3×3 convolutional layers with 128 output channels in the main path, separated by batch normalization and RELU activations. A skip connection is applied with a 1×1 convolution when the input dimensions do not match the output dimensions of the main path. The feature representation is refined with an SE block of 128 channels. The same goes for Residual Block 2, which has the same layout but 256 channels for the output on both the main path and the SE block. Lastly, Residual Block 3 has 512 output channels but follows the same architectural design.

The model utilizes a Global Pooling layer to pool spatial information. This layer is a concatenation of adaptive average pooling and adaptive max pooling outputs. The model captures complementary spatial features by combining these two pooling strategies, leading to a stronger feature representation.

The Res-SENet classifier then utilizes the pooled features through a few fully connected layers. Finally, the features are flattened and fed into a linear layer that transforms them into 512 dimensions, after which a ReLU activation is applied, followed by batch normalization and another dropout layer with a probability of 0.5. This step is followed by another linear layer that reduces to 256, then ReLU, batch normalization, and a dropout layer with p=0.3. The last linear layer accepts the features we got and maps them to the number of output classes, giving us classification logits. A softmax is then applied to the logits to get the class probabilities.

Our Spinach-ResSENet architecture was trained and run under the same, constant conditions. The AdamW optimizer was used, having the base learning rate set to and the weight decay of , and it was trained for 30 epochs using a batch size of 32. Furthermore, the Cosine Annealing with Warm Restarts scheduler was applied. In the latter, T₀ = 5, , and . In order to avoid overfitting, the early stopping technique was applied, with a patience of 5 epochs. The Cross-Entropy Loss label smoothing value was 0.1. In addition, the use of the Mixed Precision Training (AMP) technique for the improvement of the stability and the efficiency of the process was enabled. For the purpose of ensuring the reproduction of the experiments, the random state for the Python, NumPy, and Torch environments was locked to 42.

3.3.3 SpinachViT model.

The SpinachViT model, illustrated in Fig 6, is based on a ViT-Base style architecture adapted for leaf disease classification. Input images of size 224×224×3 are split into non-overlapping 16×16 patches, which are linearly projected into 768-dimensional patch embeddings. A learnable classification token (CLS) is prepended to the patch sequence, and learnable positional embeddings are added to preserve spatial information. The resulting sequence is then passed through a stack of 12 transformer encoder layers, each consisting of multi-head self-attention with 12 attention heads and a feed-forward multilayer perceptron (MLP) with expansion factor 4. Residual connections and layer normalization are applied around both the attention and MLP blocks to stabilize training.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Proposed SpinachViT architecture.

https://doi.org/10.1371/journal.pone.0340989.g006

After the final encoder layer, the CLS token is extracted, passed through layer normalization, and fed into a linear classification head that outputs logits for the three classes: Alternaria, straw mite, and healthy leaf. This configuration results in approximately 85.5M trainable parameters, providing a high-capacity transformer baseline for comparison against more lightweight CNN-based models.

The SpinachViT model was trained for 30 epochs using the AdamW optimizer (learning rate , weight decay 0.05, mini-batch size 32) with a cosine annealing learning-rate scheduler. Image data were resized to 224×224 pixels and normalized using dataset-specific statistics before patch embedding. Cross-entropy loss with dropout regularization (0.1) was used for optimization. For comprehensive evaluation, Test-Time Augmentation (TTA) was applied using eight augmented views per test image, including horizontal flips, rotations of 90^°, 180^°, and 270^°, and color jitter; predictions were averaged across views. Metrics for evaluation included accuracy, precision, recall, F1 score, confusion matrices, and macro-averaged AUC-ROC using a one-vs-rest scheme.

3.4 Self-supervised SimSiam (Vanilla ResNet-50)

SimSiam (Simple Siamese Networks) [25] was employed for self-supervised representation learning. Unlike contrastive methods that require negative pairs, very large batch sizes, or momentum encoders, SimSiam learns from positive pairs only, constructed from two independently augmented views of the same image. A stop-gradient operation stabilizes training and prevents representational collapse, yielding a simple and effective self-supervised framework.

Two SimSiam pipelines were constructed in this study: (i) a Vanilla ResNet-50 encoder and (ii) a CBAM-enhanced ResNet-50. Each model underwent (1) self-supervised pretraining using unlabeled Malabar spinach leaf images, followed by (2) supervised fine-tuning using labeled images. Fig 7 illustrates the Vanilla ResNet-50 SimSiam framework.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Proposed Vanilla ResNet-50 SimSiam pretraining and supervised fine-tuning pipeline.

https://doi.org/10.1371/journal.pone.0340989.g007

Stage I: SimSiam pretraining.

Data processing and augmentation. Unlabeled spinach leaf images were resized/cropped to pixels and normalized using ImageNet statistics. For each input image, two independent augmented views were generated using a strong augmentation pipeline:

1. RandomResizedCrop (size 224, scale 0.2–1.0)
2. RandomHorizontalFlip (p = 0.5)
3. RandomApply(ColorJitter(0.4,0.4,0.4,0.1), p = 0.8)
4. RandomGrayscale (p = 0.2)
5. GaussianBlur (kernel size 23, )
6. Tensor conversion and ImageNet normalization

This combination of spatial and photometric transformations exposes the encoder to diverse appearance variations and encourages view-invariant feature learning.

Network architecture. The encoder backbone was a ResNet-50 without its classification head (global average pooling output, 2048-D). Following the standard SimSiam design, a three-layer MLP projector was appended (Linear–BN–ReLU → Linear–BN–ReLU → Linear–BN with affine=False), along with a two-layer predictor MLP (Linear–BN–ReLU → Linear).

Objective function. SimSiam minimizes a symmetric negative cosine similarity objective:

(1)

where

(2)

Here, z is the projector output, p is the predictor output, and stopgrad blocks gradients through the target representation, preventing collapse.

Optimization. Pretraining used SGD with momentum 0.9, weight decay 10⁻⁴, batch size 64, and a base learning rate linearly scaled as . A cosine learning-rate schedule was applied for 200 epochs. Mixed precision training (torch.amp) was used for efficiency.

Stage II: Supervised fine-tuning.

Network adaptation. The SimSiam-pretrained ResNet-50 backbone weights were transferred to the supervised stage. The SimSiam projector/predictor were discarded, and a lightweight classification head was attached:

• a 512-unit fully connected layer,
• BatchNorm, ReLU, and 30% dropout,
• a final three-class linear output layer (logits).

Supervised data augmentation and evaluation protocol. Training used moderate augmentations: RandomResizedCrop (scale 0.7–1.0, ratio 0.9–1.1), RandomHorizontalFlip (p = 0.5), and RandAugment (2 operations, magnitude 7), followed by ImageNet normalization. Validation and test used a deterministic protocol: resize to 256 and center-crop to 224.

Class imbalance handling. Class imbalance was handled using an optional WeightedRandomSampler during mini-batch construction. When the sampler was enabled, Cross-Entropy was kept unweighted to avoid double-compensation; otherwise, class-weighted Cross-Entropy was used.

Optimization and training strategy. Fine-tuning ran for up to 60 epochs with early stopping (patience = 10) based on validation accuracy. AdamW was used with weight decay 10⁻⁴ and layer-wise learning rates, with a smaller learning rate for the backbone and a larger one for the classifier head. A cosine learning-rate scheduler was applied. Macro-F1 was recorded as a diagnostic metric (while early stopping was based on validation accuracy). Mixed precision training was used in both CE and hybrid settings.

Hybrid loss configuration (optional). Selected experiments used a CE-dominant hybrid objective combining Cross-Entropy and Supervised Contrastive Loss (SupCon). In the hybrid setting, two augmented views per training image were generated for SupCon, while Cross-Entropy was computed on one view. A 128-D projection head was used for contrastive embeddings in the supervised stage. The total loss was:

(3)

with and temperature . Unless otherwise stated, light mixup was enabled only in the hybrid setting () for the CE branch.

Test-Time Augmentation (TTA). For evaluation, predictions were averaged across multiple conservative test-time views using mean aggregation of softmax probabilities. The TTA policy was selected on the validation set from a small candidate set (1, 2, or 4 views) and then applied once to the test set. The 4-view policy consisted of:

• center-crop after resize 256,
• horizontal flip after resize 256,
• center-crop after resize 288,
• horizontal flip after resize 288.

Because the test set is relatively small and the TTA transforms are intentionally mild, TTA may yield performance comparable to a single forward pass; therefore, both single-crop and TTA results are reported.

Algorithm 1 Pretraining and fine-tuning of the proposed Vanilla SimSiam-ResNet-50 model.

Require: Leaf image dataset: train, validation, test, and pretrain splits

Ensure: Best-performing model, evaluation metrics, and ablation summary

1: Initialization: Load dataset splits; define ResNet-50 backbone and heads; configure augmentations, losses, and optimizers.

2: SimSiam Pretraining (Self-Supervised):

3: for each epoch do

4:   for each minibatch in pretrain set do

5:    Generate two independently augmented views per image;

6:    Forward: backbone → projector → predictor;

7:    Compute SimSiam loss using symmetric negative cosine similarity with stop-gradient;

8:    Backpropagate and update parameters using SGD;

9:   end for

10: end for

11: Save pretrained backbone weights.

12: Fine-Tuning (Supervised):

13: Discard SimSiam heads; attach classification head (and SupCon projection if hybrid);

14: Load pretrained backbone weights into supervised model;

15: Optionally enable WeightedRandomSampler for class imbalance;

16: for each epoch do

17:   for each minibatch in training set do

18:    Apply supervised augmentations (RandAugment);

19:    if Hybrid configuration then

20:     Generate two augmented views per image for SupCon; compute CE on one view;

21:    end if

22:    Compute CE (or CE + λSupCon) loss;

23:    Backpropagate and update parameters using AdamW with layer-wise learning rates;

24:   end for

25:   Evaluate on validation set (early stopping monitors validation accuracy);

26:   if validation accuracy does not improve for patience epochs then

27:    Trigger early stopping;

28:   end if

29: end for

30: Save best model weights.

31: Evaluation:

32: Load best model;

33: Compute single-crop test predictions;

34: Select a TTA policy on validation (1/2/4 views) and apply once on test;

35: Average softmax probabilities across TTA views;

36: Compute metrics: accuracy, macro ROC-AUC, confusion matrix, and diagnostic scores (e.g., macro-F1).

37: Output: Single-crop accuracy, TTA accuracy, macro ROC-AUC, ablation summary.

3.5 Channel-spatial attention SimSiam (CBAM-ResNet-50)

Stage I: SimSiam pretraining.

Pipeline: The SimSiam pretraining pipeline for the CBAM-ResNet-50 model followed the same procedure as the one for the vanilla ResNet-50 encoder. Unlabeled and non-augmented images were used, from which two strongly augmented views were generated using random resized cropping, horizontal flipping, color jittering, grayscale conversion, Gaussian blur, and ImageNet normalization. These dual views provided the positive pairs necessary for the SimSiam objective.

Network Architecture: As shown in Fig 8, the CBAM-ResNet-50 architecture extends the standard ResNet-50 by inserting a Convolutional Block Attention Module (CBAM) into every residual bottleneck block. Each CBAM block consists of:

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Proposed CBAM-ResNet-50 framework.

https://doi.org/10.1371/journal.pone.0340989.g008

Stage II: Supervised fine-tuning.

Network Adaptation: The pretrained CBAM-ResNet-50 backbone was loaded from the SimSiam checkpoint, and a classification head was used as:

• 512-unit fully connected layer
• BatchNorm, ReLU, Dropout (0.30)
• Final 3-class linear classifier

Fine-tuning incorporated AdamW optimization with layer-wise learning rates into two groups (backbone and head). For CE, the backbone learning rate was and the head learning rate was . For Hybrid, the backbone learning rate remained and the head learning rate was . A cosine annealing scheduler and early stopping were applied.

Mixup-based training was applied in the Hybrid setting (default in code), and RandAugment was used in the training pipeline to improve robustness and regularization.

Both cross-entropy and hybrid loss (CE + supervised contrastive) configurations were evaluated. In the Hybrid configuration, the objective used in code is additive:

(4)

with temperature .

This loss encourages features from the same disease class to cluster more tightly in embedding space, complementing the attention mechanisms in CBAM.

3.6 Evaluation protocol for both SimSiam frameworks

Evaluation was conducted using Test-Time Augmentation (TTA) after fine-tuning. The TTA policy was selected on the validation set from 1, 2, 4 views and then applied once on the test set. The views include center-crop evaluation at 256 resize, a horizontal-flip counterpart, and an additional resize level (288) in the 4-view setting. Final predictions were computed by averaging softmax probabilities across the selected augmented views. The reported metrics include:

• Overall accuracy
• Class-wise F1-scores
• Macro ROC-AUC
• Confusion matrix

Expected Calibration Error (ECE) was additionally reported in the robustness ablation experiments (noise training) with post-hoc temperature scaling to ensure calibrated probabilities.

3.7 Swin transformer

SwinV2, a hierarchical vision transformer architecture [26], was employed for Malabar spinach leaf disease classification. Through shifted windowed self-attention, SwinV2 constructs multi-scale feature hierarchies while maintaining linear computational complexity with respect to image size. Fig 9 illustrates the configuration incorporated in this study.

Algorithm 2 Pretraining and fine-tuning of the proposed CBAM-ResNet-50 with SimSiam.

Require: Leaf image dataset: train, validation, test, and pretrain splits

Ensure: Best-performing model, evaluation metrics, and ablation summary

Initialization:

2: Load all dataset splits (train, val, test, pretrain);

  Define CBAM-ResNet-50 backbone;

4: Configure data augmentations (strong SimSiam augs for pretrain; RandAugment for fine-tune), loss functions (CE / CE+SupCon), and optimizer.

   SimSiam Pretraining (Self-Supervised):

6: for each epoch do

    for each minibatch in pretrain set do

8:    Generate two augmented views per image;

     Pass both views through CBAM-ResNet-50 backbone and projection head;

10:    Pass outputs through prediction head;

     Compute SimSiam loss using negative cosine similarity;

12:    Backpropagate and update parameters;

    end for

14: end for

  Save pretrained weights (backbone + SimSiam heads).

16: Fine-Tuning (Supervised):

  Load pretrained backbone weights;

18: Attach classification head;

  for each epoch do

20:   for each minibatch in training set do

     Apply augmentations (RandAugment), and mixup in the Hybrid setting;

22:    Perform forward pass and compute classification loss (CE or CE+SupCon);

     Backpropagate and update parameters using AdamW;

24:   end for

    Evaluate on validation set;

26:   if validation accuracy does not improve for patience epochs then Trigger early stopping;

28:   end if

  end for

30: Save best model weights.

  Evaluation and Ablation:

32: Load best model;

  Select TTA policy on validation from 1,2,4 views and apply to test;

34: Compute metrics: accuracy, macro ROC-AUC, and confusion matrix; Perform noise ablation (Gaussian / salt-pepper / both) and report robustness; apply temperature scaling for calibrated ECE reporting.

36: Output: Final accuracy, TTA accuracy, ROC-AUC, and ablation statistics.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. Applied swin transformer architecture.

https://doi.org/10.1371/journal.pone.0340989.g009

Patch Embedding and Hierarchical Stages: Input leaf images of 256×256×3 were initially partitioned into non-overlapping 4×4 patches, generating a 64×64 grid of patch tokens. These tokens were processed through four hierarchical stages with patch merging between stages. Shifted windows were applied between successive layers, enabling cross-window information flow and facilitating global contextual reasoning over leaf surfaces. In this work, the SwinV2-Base hierarchy followed a depth configuration of [2, 2, 18, 2] with stage-wise embedding dimensions (128, 256, 512, 1024) and attention heads (4, 8, 16, 32); the window size was 12 in Stages 1–2 and 16 in Stages 3–4 (Fig 9). Patch merging halves the spatial resolution at each transition (64→32→16→8) while doubling the channel dimension.

Classification Head: Global average pooling was used to combine the final spatial feature map into a feature vector, which was then passed through a fully connected classifier to produce logits for the three disease categories: Alternaria, straw mite, and healthy leaf.

Pre-training and Optimization Objective: The SwinV2 model was implemented using the timm library. Two initialization settings were considered using the same SwinV2-Base architecture: (i) random initialization (scratch) and (ii) pretrained initialization using swinv2_base_window12to16_192to256.ms_in22k_ft_in1k. Fine-tuning used label-smoothed cross-entropy loss with smoothing factor 0.1.

Training Configuration: The controlled SwinV2-Base configuration swinv2_base_window12to16_192to256 was used with:

• Optimizer: AdamW
• Learning rate: 1×10⁻⁴
• Weight decay: 0.05
• Epochs: up to 30 (with early stopping)
• Batch size: 32
• Mixed precision training: enabled
• Early stopping: patience 7

Evaluation Protocol (Strict): To avoid center-bias from cropping, validation and testing used a strict resize protocol: all images were resized directly to 256×256 without center cropping.

Performance Overview: Under the strict evaluation protocol, the SwinV2-Base model achieved strong performance. The pretrained initialization (ms_in22k_ft_in1k) improved results compared to scratch, demonstrating the benefit of large-scale pretraining when fine-tuned on the Malabar spinach dataset.

3.8 Explainable AI (XAI)

In the study, the following Explainable AI (XAI) techniques were employed to enhance model interpretability.

• Gradient-weighted Class Activation Mapping (Grad-CAM) [27] was used to provide visual explanations for the predictions made by the applied deep learning models by highlighting class-discriminative regions in input images. The process comprised three steps: (a) Forward Pass: passing an input image through the trained network and obtaining feature maps from the target convolutional layer (usually the last conv layer before classification); (b) Gradient Computation: calculating gradients of the predicted class score concerning the feature maps by using backpropagation; and (c) Heatmap Generation: computing channel-wise importance weights by global average pooling the gradients, then linearly combining the feature maps with these weights to yield a coarse localization map allowing intuitive explanations.
• Gradient-weighted Class Activation Mapping++ (Grad-CAM++) is a newer visual explanation method that is an improvement over the regular Grad-CAM approach by including higher-order gradient information. For situations where an object or a feature needs to be highlighted more than once, this improvement allows for locating important areas in an image more accurately.
• Grad-CAM++ is an improvement over Grad-CAM by employing higher-order gradients for more precise localization of salient image regions. It weights activations based on both first and second-order derivatives, rather than mere gradient averaging, to allow more accurate highlighting of multiple discriminative regions. This technique generates sharper visual explanations without modifying the original model architecture, and it works well, especially for complicated scenes with several relevant features. The technique accomplishes this by an improved weighting scheme that captures the spatial importance of features in a better way while still being computationally light and easy to implement.
• Layer-wise Class Activation Mapping (LayerCAM) [28] was utilized to exhibit fine-grained visual explanations for the predictions of the CNN by highlighting class-relevant areas in the input images through the utilization of spatial attention obtained from activation maps. It involved a three-step process: (a) Forward Pass – forward passing an input image through the trained network and retrieving activation maps from the target convolutional layer; (b) Gradient Computation – calculating gradients of the predicted class score against the activation maps through backpropagation; and (c) Heatmap Generation – applying ReLU to the gradients and element-wise multiplication with the corresponding activation maps to retrieve spatially accurate attention maps, which were aggregated to create a fine-resolution localization map that better indicates class-discriminative areas than Grad-CAM.

4.1 Performance of Spinach-CNN

The performance of the custom SpinachCNN model in terms of training and validation on the Spinach dataset is demonstrated in Fig 10, where the trends of loss and accuracy are presented over 30 epochs. The loss curve shows a steady decrease in training and validation losses, with the training loss starting from above 1.1 and gradually decreasing to approximately 0.5. Despite some oscillations depicting slight variations in the model’s generalization capacity, the validation loss follows the same trend. The accuracy plot illustrates that the model learns effectively, as the training accuracy improves from approximately 50% to nearly 90% over the epochs. Reaching 80% validation accuracy within five epochs, the model demonstrated excellent performance and maintained a steady level of around 90%. The model is efficient, has high accuracy and good feature extraction from the Spinach dataset, and demonstrates strong generalization capability.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 10. Accuracy and Loss vs. epoch for SpinachCNN.

https://doi.org/10.1371/journal.pone.0340989.g010

4.2 Performance of Spinach-ViT

The performance of the Vision Transformer-based SpinachViT model on the employed spinach dataset is presented in Fig 11, which includes training and test loss and accuracy trends across 30 epochs. As observed from the loss plot, a smooth drop in the training loss from above 0.8 to near zero indicates good model optimization. As the test loss varies across wide ranges during training, the test loss exhibits some instability in terms of generalization. The test loss starts to increase approximately after 20 epochs, even though it initially shows a decreasing trend, potentially indicating the onset of overfitting.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 11. Accuracy and Loss vs. epoch for Spinach-Vit model.

https://doi.org/10.1371/journal.pone.0340989.g011

With training accuracy reaching close to 100% towards the final epochs, the plot for accuracy replicates significant performance. In the early stages, the test accuracy exhibits quick improvement to over 80% within the first few epochs. However, it also demonstrates significant variability, highlighting some irregularity in the ability of the model to generalize. The ViT model shows great feature learning abilities on the Spinach dataset and is highly accurate, but needs more fine-tuning to achieve optimal stability.

4.3 Performance of Spinach-Res-SENet

The performance of the Spinach Res-SENet model over 30 epochs is shown in Fig 12. Both the training loss and the validation loss decline steadily in the early epochs, as observed from the loss plot. After around 10 epochs, validation loss converges with some fluctuations, suggesting a possible convergence, but training loss continues to decline, suggesting continuous model fitting improvement.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 12. Loss and Accuracy vs. epoch of the Spinach-Res-SENet model.

https://doi.org/10.1371/journal.pone.0340989.g012

Both validation and training accuracy increase during the initial epochs. The model achieves a validation accuracy of close to 90% at the completion of training, indicating that it performs effectively. Although there are some minimal changes in the later phases, no indications of severe overfitting or underfitting are present, implying that the model capacity and regularization are well-balanced.

Fig 13 shows the classification performance of our Spinach Res-SENet model, with only negligible misclassifications. The classification for Class 1 is perfect. Class 0 has been mainly classified into Class 1 with slight misclassifications, while Class 2 has been perfectly predicted.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 13. Confusion Matrix of the Spinach-Res-SENet model.

https://doi.org/10.1371/journal.pone.0340989.g013

4.4 Performance of Spinach-SimSiam-ResNet-50

Fig 14 summarizes the test-set performance of the proposed Spinach-SimSiam-ResNet-50 with standard Cross-Entropy fine-tuning (CE). The confusion matrices show that most errors arise from confusion between Alternaria and Healthy Leaf, while straw_mite remains consistently well-separated.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 14. Test-set performance of Spinach-SimSiam-ResNet-50 with CE fine-tuning.

https://doi.org/10.1371/journal.pone.0340989.g014

Under CE fine-tuning (Fig 14 the model almost correctly classified Alternaria samples (0.95), and achieved perfect classification for both Healthy Leaf (1.00) and straw_mite (1.00). The ROC analysis further confirms near-saturated separability, with macro- and micro-average AUC values of 0.9994 and 0.9991, respectively, and per-class AUC reaching 1.0000 for straw_mite.

Under Hybrid fine-tuning, the model almost correctly classified Alternaria samples (0.92), Healthy Leaf samples (0.97), and again achieved perfect classification for straw_mite (1.00). While the Hybrid configuration increases the emphasis on feature-level class separation during training, the ROC curves show lower AUC values than CE in this particular checkpoint (macro AUC = 0.9895; micro AUC = 0.9855), with straw_mite remaining unchanged at AUC = 1.0000.

4.5 Performance of Spinach-SimSiam-CBAM-ResNet-50

Fig 15 shows the test-set performance of Spinach-SimSiam-CBAM-ResNet-50 under Hybrid fine-tuning (CE , with and ). Nearly all samples are correctly classified, with the few errors mainly occurring between Alternaria and Healthy Leaf; straw_mite remains consistently well-separated. The ROC curves further confirm strong class separability, with macro- and micro-average AUC values close to 1.0 for both objectives.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 15. Test-set performance of Spinach-SimSiam-CBAM-ResNet-50 under hybrid fine-tuning.

https://doi.org/10.1371/journal.pone.0340989.g015

4.6 Performance of swin transformer

Fig 16 evaluates the performance of the Swin Transformer model on the test dataset. The confusion matrix shows strong classification performance across all three classes. Under the strict evaluation protocol (direct resize to 256×256 with no center crop), the pretrained SwinV2-Base model achieved high accuracy and macro F1-score, confirming its robust discriminative capability for Malabar spinach leaf disease classification.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 16. Confusion matrix for the applied swin transformer.

https://doi.org/10.1371/journal.pone.0340989.g016

4.7 Comparative analysis and ablation study

Using the fixed training, validation, and test splits described in Sect 3.1, all robustness ablation experiments are conducted under the same base pipeline as the main setting (identical backbone, optimizer schedule, and training budget). The only controlled variable is the injection of synthetic noise into the training images only through the preprocessing/augmentation pipeline. Evaluation is always performed on the original clean validation and test sets.

Four noise conditions are studied:

• No Noise: Standard augmentation without additional noise.
• Gaussian Only: Additive Gaussian noise is applied to every training image.
• Salt-and-Pepper Only: Salt-and-pepper noise is applied to every training image.
• Both Noises: Gaussian noise and salt-and-pepper noise are applied sequentially to each training image.

As illustrated in Fig 17, noise augmentation is applied exclusively during training to assess whether the model can learn robust representations under perturbations while maintaining generalization to clean inputs. All experiments are performed on the SimSiam-CBAM-ResNet-50 architecture, initialized from the same SimSiam in-domain pretraining (Sect 3.5) and fine-tuned using the cross-entropy objective (Sect 3.4). For fairness and reproducibility, we reuse the best-performing CE seed identified in our main three-seed evaluation and keep the remaining hyperparameters unchanged.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 17. Examples of noise-augmented training images used in the robustness ablation study.

https://doi.org/10.1371/journal.pone.0340989.g017

Evaluation metrics.

We report Test Accuracy, Macro ROC-AUC, and Expected Calibration Error (ECE). Since calibration can be significantly improved post hoc without affecting classification accuracy, we report temperature-scaled ECE (ECE(TS)) computed by fitting a single temperature parameter on the validation logits and evaluating calibration on the test set.

Table 2 shows that SimSiam-CBAM-ResNet-50 maintains strong discriminative performance under synthetic perturbations introduced during training. While both single-noise settings produce only a modest reduction in accuracy relative to the clean-training baseline, the combined noise setting yields the largest degradation, indicating additive difficulty when multiple corruption types are applied concurrently. Macro ROC-AUC remains high across all conditions (), suggesting that class-agnostic separability is preserved even when the training distribution is corrupted. In terms of reliability, temperature scaling substantially improves calibration, keeping ECE(TS) within a moderate range (3.31–6.36%), even for the most challenging combined-noise condition.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Ablation study: robustness of SimSiam-CBAM-ResNet-50 to synthetic noise (training-only noise).

https://doi.org/10.1371/journal.pone.0340989.t002

4.8 Visualizations and qualitative results

XAI techniques were applied to the applied SimSiam-CBAM-ResNet-50 model, which demonstrated excellent classification accuracy on the test set, with an average confidence score of around 94.4% (±3.8%). Class-wise analysis revealed that predictions for straw mite exhibited the highest mean confidence, about 96.3%±2.5%, followed by healthy 95.3%±3.6%, and Alternaria 90.1%±2.4%. Analysis using Grad-CAM, Grad-CAM++, and Layer-CAM consistently highlighted biologically important regions in all samples, with Grad-CAM++ showing superior localization (mean intensity: 0.75) compared to Grad-CAM (0.72) and Layer-CAM (0.68). Fig 18 compares CAM visualizations for a straw mite-infected leaf (96.4% confidence). All methods highlight mite damage patterns, with Grad-CAM++ showing sharper lesion localization than Grad-CAM’s broader activations, while Layer-CAM reveals finer details. The model’s decision was validated by the consistent focus on biologically relevant features.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 18. XAI Explainability Heatmap with Grad-CAM, Grad-CAM++, and LayerCAM.

https://doi.org/10.1371/journal.pone.0340989.g018

According to Table 3, the comparative performance shows several salient findings across model architectures. The Spinach-CNN baseline recorded test accuracy of 91.00% and TTA accuracy of 96.68% using 5.49M parameters, indicating that even a lightweight CNN can benefit from test-time augmentation but remains limited in single-crop generalization. Spinach-ResSENet performed best among the compact ResNet-based variants, achieving 96.01% test accuracy and 95.35% TTA accuracy with 5.53M parameters, showing that squeeze-and-excitation (SE) attention improves feature quality without increasing model size substantially. The Spinach-ViT model, trained from scratch, produced comparatively lower scores (90.7% test, 91.3% TTA) despite its larger capacity (85.5M parameters), which is consistent with transformer training being more data-hungry under limited in-domain supervision. Unless otherwise specified, confusion matrices and example ROC curves in this section are plotted for the best-performing random seed of each configuration, whereas Table 3 reports mean ± standard deviation over three seeds. This explains small numerical differences between some per-class results in the figures and the aggregate statistics in the table.

For the self-supervised pipelines, SimSiam-ResNet-50 shows improved results over the simple CNN baseline when averaged over seeds. Under CE fine-tuning, SimSiam-ResNet-50 achieved % test accuracy with macro ROC-AUC . The Hybrid objective (CE SupCon) increased mean accuracy to %, although its macro ROC-AUC in this protocol was . After incorporating channel-spatial attention, SimSiam-CBAM-ResNet-50 (CE) delivered the strongest overall performance among the SimSiam-based models, reaching % test accuracy with identical TTA accuracy (%) and a near-saturated macro ROC-AUC of at M parameters. This highlights that CBAM attention provides a consistent gain when paired with in-domain SimSiam pretraining and standard CE fine-tuning.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Performance comparison of self-supervised, attention-based, and transformer-based models.

https://doi.org/10.1371/journal.pone.0340989.t003

The SimSiam-CBAM-ResNet-50 (Hybrid) configuration shows a modest reduction in accuracy compared to its CE counterpart, achieving % test accuracy (TTA %) with macro ROC-AUC . Although supervised contrastive learning (SupCon) often improves generalization, its effect depends strongly on dataset size, class structure, and batch composition. In our setting, the hybrid objective can reasonably yield lower accuracy for several reasons. First, the hybrid loss optimizes a representation geometry rather than the evaluation metric. The CE term drives the classifier toward decision-boundary separation, whereas SupCon encourages tighter intra-class clustering and larger inter-class angular margins. For fine-grained leaf disease imagery, where Alternaria and healthy leaves may share subtle texture cues, excessive compactness can unintentionally suppress class-discriminative variations that CE alone would preserve. Second, the small dataset magnifies gradient variance. SupCon relies on multiple same-class positive pairs per batch, but with moderate mini-batch sizes and class-balanced sampling, the number of positives per class can be low and inconsistent. This increases optimization noise and can create gradient conflict between CE and SupCon, especially when the CE-only model is already operating near ceiling performance. Consequently, while the hybrid loss improves representation structure (as reflected in high ROC-AUC), it may act as mild over-regularization during fine-tuning, leading to the observed small decrease in top-1 accuracy.

Finally, the transformer-based models achieved the strongest overall results in this evaluation setting. SwinV2-Base trained from scratch under strict evaluation obtained 94.95% test accuracy and 95.96% TTA accuracy with macro ROC-AUC 0.9970 at 86.90M parameters. The pretrained SwinV2-Base (ms_in22k_ft_in1k) achieved the highest performance overall, reaching 97.98% test accuracy and 98.99% TTA accuracy with a perfect macro ROC-AUC of 1.0000, again at 86.90M parameters. Overall, the results show that attention-augmented self-supervised learning (SimSiam-CBAM) provides a strong parameter-efficient alternative to large transformers, while the pretrained Swin model delivers the best absolute accuracy at a substantially higher parameter cost.