2.1 Dataset acquisition

Black pod disease is a major fungal infection that affects cocoa (Theobroma cacao) worldwide, caused mainly by species of the genus Phytophthora. The primary culprits include P. palmivora, P. megakarya, and P. capsici, whose spores spread rapidly during heavy tropical rains. Early symptoms appear as small yellow spots on cocoa pods, which quickly turn brown and expand to cover the entire pod within about five days. Infected pods often develop a distinctive white mycelial growth, a clear sign of the pathogen’s presence. Frosty pod rot, caused by the basidiomycete fungus Moniliophthora roreri, is a devastating disease of cocoa (Theobroma cacao) worldwide. Alongside witches’ broom disease caused by Moniliophthora perniciosa, it poses a major threat to cacao production. Early symptoms include small, water-soaked lesions on pods that enlarge and become necrotic. As the disease advances, pods develop a thick, white, powdery growth, later shriveling, hardening, and mummifying while still attached to the tree.

The dataset for this study was developed with approval from the Ghana Cocoa Board, Twifo Praso district. The Cocoa pods were photographed with Canon 60D and Samsung Galaxy S22 in a cocoa plantation located at Wawasi, a village a few km from Twifo Praso in the central region of Ghana (5°36’59.99“ N -1°32’59.99” W). Images of a single class were taken so that the folders could be more easily classified and the camera was maintained at a distance of approximately 50 cm from the pod, ensuring that no shadows were cast on the pods due to sunlight. The photographs were captured at various periods of the day, in the morning between 6:00am and 11:00am, in the afternoon from 12:00 pm to 2:30 pm and in the evening between 3:00 pm and 5:00 pm, from different angles in an uncontrolled environment. The dataset was created between 20th January, 2023 and March 5th 2024, the main season for cocoa harvesting in the research region.

All images were captured at a native resolution of 5184 × 3456 × 3 pixels (RGB). Images underwent quality control review, and those with significant blur, occlusion, or poor lighting conditions were excluded. The uniform native resolution eliminated the need for extensive preprocessing related to resolution variability, though minor cropping was performed during manual inspection to remove irrelevant background elements

The images were classified into three categories: Phytophthora-infected (Phyto), Moniliophthora-infected (Moni), and Healthy cocoa pods. The annotation process was conducted by three independent expert annotators, including two plant pathologists from the Ghana Cocoa Board with a combined 15 years of experience in cocoa disease identification and one senior agronomist specializing in cocoa cultivation. The framework for detecting cocoa pod diseases strictly guided the selection process [42,80]. Each annotator independently labeled all 1,704 images based on visual symptoms following established diagnostic criteria: Phyto infection was identified by characteristic brown necrotic lesions with white mycelial growth; Moni infection by powdery white growth and pod mummification; Healthy pods showed no visible disease symptoms and maintained normal coloration. Following initial independent annotation, a consensus meeting was held to resolve discrepancies. Images with disagreement were re-examined collectively, and final labels were assigned based on majority agreement and discussion of diagnostic features. All images were manually checked to remove irrelevant and duplicate images. Importantly, all annotation was performed on the original high-resolution images of 5184 × 3456 × 3 pixels

To assess annotation reliability, Cohen’s kappa coefficient was calculated for pairwise agreement among the three annotators [81]. The inter-annotator agreement scores were: Annotator 1 vs. Annotator 2 (κ = 0.91), Annotator 1 vs. Annotator 3 (κ = 0.89), and Annotator 2 vs. Annotator 3 (κ = 0.93), indicating excellent agreement (κ > 0.80). Overall Fleiss’ kappa across all three annotators was 0.91, confirming high reliability in disease classification. Of the 1,704 images, 152 (8.9%) required consensus discussion, with disagreements primarily occurring in early-stage infections where disease symptoms were subtle

The final dataset comprises 1,704 images divided into three balanced categories: Phyto (568), Moni (568), and Healthy (568), each representing exactly one-third of the dataset. Table 1 summarizes the class distribution. The balanced distribution ensures the dataset is suitable for training machine learning models without class imbalance issues, preventing bias toward any particular category during model training and evaluation.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Summary data of split disease conditions.

https://doi.org/10.1371/journal.pone.0348147.t001

Each image has a uniform resolution of 5184 × 3456 × 3 pixels (RGB color space), providing high-quality detail necessary for accurate disease feature extraction. The Phyto class displays dark brown to black necrotic areas with clear demarcation from healthy tissue and occasional white fungal growth. The Moni class exhibits white powdery spore masses covering substantial pod portions, deformed shapes, and signs of premature hardening. The Healthy class shows uniform green, yellow, or orange coloration (depending on maturity stage), smooth texture without lesions, and absence of fungal growth or discoloration.

Fig 1 illustrates representative samples from each class, demonstrating visual quality and diversity captured under various lighting conditions and angles. The examples show the distinct visual patterns that differentiate each disease category. These visual examples confirm the dataset’s suitability for training discriminative models capable of distinguishing between disease states based on observable phenotypic features.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Samples of dataset classes.

https://doi.org/10.1371/journal.pone.0348147.g001

The Cocoa Disease GH dataset was compared with four benchmark datasets: Cocoa Diseases (YOLOv4), Cacao Diseases in Davao, Black and Borer Pod Rot, and Coffee and Cocoa datasets. Table 2 Provides a detailed description of all datasets, including sample sizes, class distributions, and image acquisition conditions.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Summary of datasets used in the study.

https://doi.org/10.1371/journal.pone.0348147.t002

The Cocoa Diseases (YOLOv4) dataset, created in [82] and hosted on Kaggle, contains images with a resolution of 3120 × 4160 pixels for classification and object detection tasks. It is organized into three categories: 100 healthy pods (“Sana”), 107 Black Pod Rot pods (“Fito”), and 105 Frosty Pod Rot pods (“Monilia”). The Cacao Diseases in Davao dataset, developed in [83], contains about 4,300 images at 1080 × 1080 pixels, categorized into Healthy, Black Pod Rot, and Pod Borer classes, with 3,344 healthy pods, 943 Black Pod Rot images, and 103 Pod Borer images. The Black and Borer Pod Rot dataset, created in [84] and available on Kaggle, contains 2,436 cocoa pod images for training and validation, with resolutions ranging from 1080 × 1080–2160 × 2160 pixels, and includes 3,595 normal pods, 906 Black Pod Rot pods, and 188 Pod Borer pods. The Coffee and Cocoa dataset, hosted on Roboflow Universe by [85], contains 3,806 images, including 3,114 cocoa pod images in JPG format with a resolution of 608 × 608 pixels. These cocoa pod images are categorized into Black Pod Rot (823 images), Frosty Pod (758 images), Mirid Pods (740 images), and Normal Pods (793 images). Together, these datasets provide a diverse and detailed range of cocoa pod images for use in classification and object detection research.

2.2 Data preprocessing

The researchers preprocessed cocoa pod images by resizing them to 224 × 224 pixels and applied extensive data augmentation techniques to expand their initial dataset of 1,704 training samples to 73,500 samples through geometric transformations (rotations of −45° to +45° and 90°, horizontal and vertical flipping, and scaling within 0.8-1.2x range), lighting adjustments (±20% brightness variation), and color modifications (±10° hue shifts) to create a more robust training dataset that would help their machine learning model accurately identify phyto-infected, moni-infected, and healthy cocoa pods under various real-world conditions including different orientations, sizes, lighting environments, and growth stages.

2.3 Feature extraction

In deep learning, feature extraction is the process of automatically determining what the best features of unprocessed data are, often using deep neural networks. Deep learning models are able to discover hierarchical feature representations directly through the data during training, as opposed to traditional machine learning approaches which may require human-designed features [13]. This process transforms complex, high-dimensional data into more meaningful and manageable representations that are necessary for downstream tasks such as classification.

2.4 Attention module

The channel attention is achieved by training a channel-wise attention map to highlight significant channels. It applies GAP and GMP in the global context with respect to spatial dimension. The input feature map is subjected to GAP and GMP operations and generates two statistics per channel [86]. The statistics are fed through a common MLP whose only hidden layer is ReLU. This leads to the final result of a channel attention vector that will be applied to the original values of the feature map for recalibration. This attention mechanism selects the important information in the picture with an exclusion of the irrelevant information. The first processing occurs on an input feature in parallel with both average-pooling and max-pooling as illustrated in the channel module in Fig 2 below.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Channel attention module.

https://doi.org/10.1371/journal.pone.0348147.g002

The multilayer perceptron (MLP) then uses a single hidden layer to transmit these two kinds of data. Finally, utilizing component-based aggregation, the resulting features are aggregated. Channel Attention is expressed as follows:

(1)

Where performs channel-wise average pooling, compressing spatial dimensions. performs channel-wise max pooling for complementary information. Two separate MLPs process these pooled features. The results are summed and passed through a sigmoid activation σ. The final channel attention is given by the formula:

(2)

The spatial attention module highlights important spatial locations for each channel by concentrating on spatial zones of interest. The spatial attention map is generated through GAP and GMP applied on the channel dimension that generates two feature maps. These maps are combined to generate a spatial attention map, which is subjected to a two-dimensional convolution and sigmoid activation. The input feature map is spatially refined using this map. After channel attention and spatial attention operations, the output feature map, also called the attention map, is adjusted by multiplying with the channel and spatial attention maps. The recalibrated feature map, which has been improved both geographically and channel-wise by the attention mechanisms, is the end result, as shown in Fig 3.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Spatial attention.

https://doi.org/10.1371/journal.pone.0348147.g003

Spatial attention attempts to find out the most important area of the input data feature after the channel attention module processes the features. The output features are processed in parallel through average-pooling and max-pooling, followed by a convolutional layer process. Spatial attention would be as follows;

(3)

Where performs spatial-wise average pooling across channels. performs spatial-wise max pooling across channels. A 7 x 7 convolution is applied to process these concatenated elements. The calculation of the spatial attention map is passed to sigmoid activation σ. The resulting spatial attention is provided by formula:

(4)

By progressively integrating channel and spatial attention, CBAM leverages both cross-channel and spatial relationships of features. More specifically, it emphasizes useful channels and strengthens local regions that are informative. The model can focus on important spatial regions and informative channels thanks to this dual attention process, which enhances the network’s representational capabilities.

Attention Module in CBAM conjunctionally loops the outputs and inputs for channel attention and spatial attention, respectively [87]. CBAM can use both spatial and cross-channel relationships of information so that information can be sculpted in a specific way to give the network instructions on what and where to concentrate by successively applying channel attention and spatial attention, as evidenced by Fig 4. To define further, it focuses on important channels and supplements useful regions. Two of the ways that the CBAM aggregates the spatial data are global maximum pooling and global average pooling. Combining the two pools can guarantee thorough extraction of high-level features and remove redundant data, making it possible to accurately learn how several channels are interdependent.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Channel, spatial attention (CBAM).

https://doi.org/10.1371/journal.pone.0348147.g004

2.5 Proposed learnable gated fusion CBAM (LGF-CBAM)

To further improve the representational capability of convolutional neural networks, we introduce a new attention mechanism, Learnable Gated Fusion CBAM (LGF-CBAM). This module is an extension of the original CBAM and adds a learnable fusion mechanism between the channel and spatial attention pathways. The CAM focusses on ‘what’ to emphasize by computing an attention map Mc(F), which is applied to the input feature map F to produce Fch = Mc(F)⊗F where ⊗ denotes element-wise multiplication. The SAM highlights the ‘where’ to be highlighted, generating a spatial attention map Ms(F), that yields Fsp = Ms(F) ⊗ F. Instead of simply adding Fch and Fsp as the two outputs, LGF-CBAM uses a gated fusion mechanism governed by learnable weights α and β, constrained by a softmax function to ensure α + β = 1. These gates are dynamically computed from global descriptors of Fch and Fsp using a lightweight MLP, such that;

(6)

Where GAP denotes Global Average Pooling

The final output is obtained as a weighted combination of both attention-enhanced maps using the formula;

(7)

This learnable gating mechanism as seen in Fig 5 allows the network to adaptively prioritise spatial or channel cues based on context, thus enhancing feature discrimination and improving task-specific performance.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Learnable gated fusion CBAM (LGF-CBAM).

https://doi.org/10.1371/journal.pone.0348147.g005

The Learnable Fusion Weights is achieved using [α, β] = Softmax(MLP([g_ch, g_sp]))

which adaptively learns to favor channel or spatial attention based on feature context. In other words, instead of the simple addition of Fch and Fsp, the fusion is learnable using the gating parameters α and β derived from a softmax operation. The Softmax normalization ensures interpretability and stability of weight assignment. The MLP learns the importance of each attention path and outputs logits, and the parallel pathways enhance flexibility compared to strict sequential attention. The final fused attention map F_attn is obtained as a weighted sum of both attention outputs. This formulation allows the model to adaptively prioritize spatial or channel information during training.

This novel architecture provides a more sophisticated attention mechanism that maintains computational efficiency while offering enhanced representational capabilities through a learnable fusion of parallel attention pathways. Unlike traditional sequential CBAM, both attention modules process the original feature map F independently, allowing for truly parallel computation and avoiding potential information loss from sequential processing. The LGF mechanism enables each stage to dynamically balance channel versus spatial attention based on the complexity and characteristics of features at that level.The Pseudocode for the LGF-CBAM Module is as follows;

Input: Feature map F

1. Compute channel attention:
   1. Apply global average pooling and max pooling on F. F^avg_c = GAP(F), F^max_c = GMP(F)
   2. Pass both through shared MLP. M^avg = MLP(F^avg_c), M^max = MLP(F^max_c)
   3. Sum outputs and apply sigmoid to obtain M_c(F). M_c(F) = σ(M^avg + M^max)
   4. Multiply M_c(F) with F to obtain F_ch.

(8)

1. Compute spatial attention:
   1. Apply average pooling and max pooling across channel dimension of F_ch. F^avg_s = AvgPool_c(F_ch), F^max_s = MaxPool_c(F_ch)
   2. Concatenate pooled maps. F^cat_s = [F^avg_s; F^max_s]
   3. Apply 7 × 7 convolution and sigmoid to obtain M_s(F). M_s(F) = σ(Conv_7 × 7(F^cat_s))
   4. Multiply M_s(F) with F_ch to obtain F_sp.

(9)

1. Compute learnable gates:
   1. Apply global average pooling on F_ch and F_sp. g_ch = GAP(F_ch), g_sp = GAP(F_sp)
   2. Concatenate descriptors and pass through MLP. MLP([g_ch, g_sp])
   3. Apply softmax to obtain α and β such that α + β = 1.

(10)

1. Fuse attention maps:

(11)

2.6 ResNet backbone with attention

Higher deterioration and more saturated accuracy are associated with deeper architectures. Deep layers in deeper networks are unable to accommodate the desired underlying mapping needed to send the result to the output. In simple terms, the more layers there are, the greater the training or test errors. With very deep frameworks, this may cause learning to slow down or learning may stall completely. Quite remarkably, and perhaps counterintuitively, overfitting is not why this decline happens, as a deep enough model does hurt its training error as outlined and verified in depth during the cited works analysis. The decreasing trend in training accuracy shows that not all systems can be optimized that easily. Fig 6 illustrates the proposed CNN model.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. The proposed CNN model.

https://doi.org/10.1371/journal.pone.0348147.g006

ResNet-101 holds the backbone of the study, that is, 49 convolutional layers and one fully connected layer. In order to address the issue of degradation, ResNet introduced skip connections, sometimes referred to as residual connections or residual learning, which enable the gradient to bypass specific network layers. Similar to the other connections, the shortcut connections do not render the network more computationally complex because they enact identity mapping and combine their outputs with the outputs of the stacked layers. This is such that the output of a given layer is added to or fed into the adjacent layer without passing through it. The model’s complexity does not significantly rise with the addition of such connections; rather, it improves the convergence of deeper models and helps them optimize better than models without residual connections. Again, it provided better performance and more effective learning, particularly in very deep systems like ours.

2.7 Parameter selection for training the proposed model

The proposed model was developed using Keras 2.4.3 with Python 3.7 and a TensorFlow 2.2.1 backend. Training was executed on Google Colab using an A1000 GPU. The performance of the model was systematically compared against current state-of-the-art networks.

The Cocoa Disease GH dataset was divided into 70% training, 20% validation, and 10% testing sets using the ShuffleSplit function in scikit-learn version 0.23.2. Because the dataset was relatively small, data augmentation techniques such as rotation, color adjustments, and horizontal and vertical flipping were applied. These steps increased sample diversity, reduced class imbalance, and helped the model learn more robust features while minimizing overfitting.To determine the most suitable training settings, ablation studies were carried out on key hyperparameters. The results of these experiments are summarized in Tables 3 and 4,.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Ablation study for optimizer selection.

https://doi.org/10.1371/journal.pone.0348147.t003

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Ablation study for activation function selection.

https://doi.org/10.1371/journal.pone.0348147.t004

Among the optimizers tested, Adam consistently delivered the best performance. On the Cocoa Disease GH dataset, it achieved the highest classification accuracy of 98.95% and the lowest validation loss of 0.0079, outperforming RMSProp, Adagrad, SGD with momentum, and standard SGD. Although SGD-based methods converged slightly faster within 72–76 epochs and required marginally less training time, they showed weaker generalization, reflected in higher validation losses. Adam reached convergence at epoch 82 with a training time of 430 minutes, demonstrating that the slightly longer training period led to better feature learning and predictive accuracy. The small gap between training and validation accuracy of 99.96% against 98.95% indicates minimal overfitting. Similar patterns were observed across the other benchmark datasets, where accuracy ranged from 94.00% to 98.95%, confirming the reliability of the optimizer.

ReLU was selected as the activation function because of its stable learning behavior and strong classification performance. It produced the same peak accuracy of 98.95% and the lowest validation loss. With a true positive rate of 99.10% and a false positive rate of just 0.89%, the model demonstrated reliable detection with few misclassifications. While Leaky ReLU performed comparably and trained slightly faster, ReLU provided more consistent gradient flow and smoother training. Its simple operation, max(0, x), also reduced computational cost. In contrast, Sigmoid and Tanh showed lower accuracy due to vanishing gradient issues.

Different batch sizes were also tested to understand their impact on performance. As shown in Fig 7, accuracy improved as the batch size increased to 16, reaching 98.60% due to more stable gradient updates. The best result, 98.95%, was obtained with a batch size of 32, which offered a good balance between stability and generalization. Increasing the batch size to 64 slightly reduced accuracy to 98.40%, suggesting weaker generalization.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Ablation study for batch size selection.

https://doi.org/10.1371/journal.pone.0348147.g007

The model was further evaluated with various pooling strategies to identify the most effective way to reduce feature map size without losing important disease-related information. Results in Fig 8 show that Global Average Pooling (GAP) performed best, achieving 98.95% accuracy by preserving meaningful features while reducing noise and dimensionality. L2-Norm pooling also showed strong performance. Global Max Pooling captured dominant features but missed finer details, while Stochastic and Median pooling introduced variability or lost important information. Global Min Pooling performed worst, indicating that weak activations were not helpful for classification. Overall, pooling methods that summarized broader feature information were more effective.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Ablation study for pooling operation selection.

https://doi.org/10.1371/journal.pone.0348147.g008

A CBAM reduction ratio of 16 was chosen to balance attention strength with computational efficiency. A learning rate of 0.0001, determined through grid search, ensured stable convergence, while early stopping halted training automatically at epoch 82 using the best model weights. Using categorical cross-entropy and standard evaluation metrics, the final model showed strong generalization with only a 1.01% difference between training and validation accuracy, confirming that the selected hyperparameters were appropriate and effective as presented in Table 5.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Selected hyperparameters for training the model.

https://doi.org/10.1371/journal.pone.0348147.t005

3.1 Feature extraction

Feature extraction time was measured as the time required for each CNN model to compute features from input images up to the final convolutional or pooling layer before classification. Timing was recorded using a batch size of 32 images on a Google Colab NVIDIA A100 GPU. The results are presented in Table 6.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Summary of feature extraction time for base models.

https://doi.org/10.1371/journal.pone.0348147.t006

In comparing backbone networks, clear differences emerged between VGG and ResNet families. VGG models produce feature representations that are roughly twice as large as those from ResNet-based models, which increases memory use and can affect downstream efficiency. ResNet and ResNetV2, by contrast, generate compact 2048 dimensional features through convolutional pooling. Although deeper ResNet variants take longer to process, they tend to learn more reliable patterns. ResNetV2 further improves training stability through batch normalization and pre-activation before convolution, allowing smoother gradient flow and more efficient learning.

3.1.1 Training time and inference time comparison of pretrained networks.

Training time was measured as the total time required for each model to complete training on the dataset, while inference time was recorded as the time required to classify a new image after training. Measurements were taken on the Google Colab NVIDIA A100 GPU. The results are presented in Table 7 whivh presents the training and inference times recorded for the baseline models

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Training and inference time summary for base models.

https://doi.org/10.1371/journal.pone.0348147.t007

VGG-16 and VGG-19 use simple, consistent designs with moderate parameter counts, enabling fast training and reliable feature extraction. Their straightforward architecture supports quicker optimization, making them suitable for rapid prototyping and scenarios with limited computational resources. ResNet models introduce a trade-off between depth and computational demand. As depth increases from ResNet-50 to ResNet-152, parameters rise significantly, extending training time and resource requirements. While deeper networks capture more complex patterns, skip connections add computational overhead despite improving training stability. For example, training time can increase from hours to nearly a full day as depth grows. ResNetV2 addresses some of these efficiency challenges through pre-activation blocks that improve gradient flow and training efficiency. Overall, VGG models offer speed and simplicity, whereas ResNet variants deliver higher accuracy when sufficient computational resources are available, with ResNetV2 providing a balance between performance and efficiency.

3.2 Classification

In order to thoroughly learn about the performance of a model, we must subject it to a number of performance metrics. In this research, the base models have been measured on regular methods of evaluation like accuracy, precision, recall or sensitivity, F1 score, and fallout. Each of these measures draws attention to the various dimensions of the model performance, particularly, in those cases when the data could not be equally distributed among categories.

3.2.1 Pretrained network.

To see how well the pretrained models could tell the classes apart, evaluation was done using metrics like precision, recall, F1 score, and false positive rate. ResNet101 was chosen as the base model based on its performance on the various metrics

As it can be seen in the Fig 9 above, ResNetV2-101 achieves the highest accuracy at 87.95%, confirming the superiority of the ResNetV2 architecture. ResNetV2-152 and ResNetV2-50 also perform strongly, with accuracies above 86%. In contrast, VGG-19 and VGG-16 show lower accuracy, especially VGG-16 at 80.75%. This indicates that VGG models are less suitable for high-precision tasks. Overall, ResNetV2 models clearly outperform their counterparts in classification performance. ResNetV2-101 achieves the lowest FPR of 0.078 and the highest TPR of 0.922, making it the most reliable model for minimizing false alarms while correctly detecting positives. This is an indicator of its high ranking performance. Contrarily, VGG-16 has the overall highest FPR of 0.150 and the lowest TPR of 0.850, which means that there is a greater likelihood of missing detections and false positives. This reduces the applicability of the VGG models in high stakes, or sensitive classification problems.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. a: Accuracy for Pretrained Models. b. Model FPR for Pretrained Models. c. Model F1 score for Pretrained Models. d. Model PPV score for Pretrained Models. e. Model TPR score for Pretrained Models. f. Model Loss Score for Pretrained Models.

https://doi.org/10.1371/journal.pone.0348147.g009

This trend is reflected in the F1 score and predictive precision.

ResNetV2-101 achieved the highest F1 score (86.70%) and PPV (88.00%), indicating a good balance between precision and recall. VGG-16 lagged behind with an F1 score of 78.40%. Model loss values further confirmed this pattern: ResNetV2-101 had the lowest loss (0.276), while VGG-16 had the highest (0.415). Based on these results, ResNetV2-101 was selected as the base model for the proposed system (Table 8).

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. Summary of evaluation on the base pretrained CNN models.

https://doi.org/10.1371/journal.pone.0348147.t008

To statistically validate the superior performance of ResNetV2-101, paired t-tests were conducted comparing its accuracy against all other base pretrained CNN models across five independent runs. Table 9 summarizes the mean accuracy differences, standard deviations, t-values, and p-values. All comparisons yielded p-values below 0.05, confirming that the performance improvements of ResNetV2-101 over ResNet-50, ResNet-101, ResNet-152, ResNetV2-50, ResNetV2-152, VGG-16, and VGG-19 are statistically significant. Standard deviation values represent the variability of the pairwise differences across the five runs, calculated as SD = Mean Difference/ (t-value/ √n), where n = 5. These results provide strong evidence that ResNetV2-101’s higher accuracy is not due to random chance and reliably outperforms alternative base models

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Statistical significance test on the base pretrained CNN models.

https://doi.org/10.1371/journal.pone.0348147.t009

3.3 Proposed model

Figs 10 and 11 present the training and validation accuracy and loss of the proposed model on the Cocoa Disease GH dataset. Training loss decreased steadily from 0.89 to 0.23 by epoch 30 and approached zero by epoch 82, while training accuracy reached 99.96%. Validation accuracy improved from 41.0% at epoch 1 to 98.95% at epoch 82, where EarlyStopping halted training to prevent overfitting.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. Training and validation accuracy for the cocoa disease GH dataset.

https://doi.org/10.1371/journal.pone.0348147.g010

Download:

• PNG
  larger image
• TIFF
  original image

Fig 11. Training and validation loss on the cocoa disease GH dataset.

https://doi.org/10.1371/journal.pone.0348147.g011

To ensure stable and efficient optimization, we adopted a two-phase monotonically decaying learning-rate schedule. A base learning rate of 1 × 10 ⁻ ⁴ was first selected via grid search as the most stable value. Training began with a higher rate of 5.1 × 10 ⁻ ³ to accelerate early learning and was gradually reduced each epoch until reaching the base rate by epoch 30.

To reduce overfitting, several regularization mechanisms inherent to the network design and training strategy were employed. L2 weight regularization (λ = 1 × 10 ⁻ ⁴) constrained large parameters and promoted simpler models, while data augmentation, including random flips, ± 15° rotations, 10–20% zoom, and ±20% brightness variations, increased training diversity and reduced memorization. Batch normalization layers embedded within each ResNetV2 block stabilized activations and provided implicit regularization, and global average pooling replaced large fully connected layers, reducing the classification head to only 6,147 trainable parameters. The decayed learning-rate schedule further acted as implicit regularization, and early stopping with a patience of 10 epochs halted training at epoch 82 once validation performance plateaued, resulting in a small generalization gap of only 0.60%.

Training curves (Figs 7 and 8) show smooth convergence. Training accuracy rose from 57% to 99.98%, while validation accuracy improved from 41% to 99.10%. Loss values steadily decreased and remained closely aligned, indicating stable learning with minimal overfitting.

The confusion matrix in Fig 12 summarizes the classification performance of the proposed model across the three categories: Phyto, Healthy, and Moni. For the Phyto class, 1,134 samples were correctly identified, while 7 and 12 samples were misclassified as Moni and Healthy, respectively. The Healthy class achieved perfect classification with no mispredictions, indicating excellent separability between healthy and diseased samples.For the Moni class, 1,143 samples were correctly classified, with 7 misclassified as Phyto and 3 as Healthy. Overall, the model demonstrates high classification accuracy across all categories. Most errors occurred between the two disease classes, with 12 Phyto samples predicted as Moni and 7 Moni samples predicted as Phyto, yielding 19 inter-disease confusions. This trend suggests that distinguishing between similar infections is more challenging than separating healthy and diseased pods. Both diseases exhibit comparable visual characteristics, including necrotic lesions, discoloration, and surface texture degradation, particularly during early infection stages when symptoms are subtle.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 12. Confusion matrix for the test dataset on the cocoa disease GH dataset.

https://doi.org/10.1371/journal.pone.0348147.g012

On the Cocoa_Disease_Gh dataset, the model achieved a TPR of 99.10% and a low FPR of 0.89%, with F1 and PPV values also around 99%. Similar performance was observed on the Cocoa Diseases (YOLOv4) dataset with 98.53% accuracy. Strong results were also recorded on the Black and Borer Pod Rot dataset (97.96%). Performance dropped slightly on the Cacao Diseases in Davao dataset at 96.19% and more noticeably on the Coffee and Cocoa dataset at 94.00%, which showed higher variability and noise (Table 10).

Download:

• PNG
  larger image
• TIFF
  original image

Table 10. Summary of model performance metrics on the cocoa disease GH dataset.

https://doi.org/10.1371/journal.pone.0348147.t010

To improve model interpretability, Grad-CAM visualizations were generated from the final convolutional layer of the proposed LGF-CBAM model, as shown in Fig 13. For clarity, each sample is presented as a triplet comprising the original image, the attention heatmap, and the heatmap overlaid on the original image. The resulting attention maps show that the network focuses primarily on lesion and infected regions of the cocoa pod while suppressing background elements, indicating that predictions are driven by disease-specific features rather than spurious artifacts.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 13. a: Grad-CAM visualizations of Healthy Cocoa Disease.b: Grad-CAM visualizations of Moni Cocoa Disease. c: Grad-CAM visualizations of Phyto Cocoa Disease.

https://doi.org/10.1371/journal.pone.0348147.g013

Table 11 presents the ablation comparison between the proposed LGF-CBAM architecture and several CBAM-integrated backbone networks to evaluate the contribution of the learnable gated fusion mechanism to classification performance.

Download:

• PNG
  larger image
• TIFF
  original image

Table 11. Performance comparison of the proposed LGF-CBAM model with baseline CBAM-based architectures on the Cocoa Disease GH dataset.

https://doi.org/10.1371/journal.pone.0348147.t011

The results in Table 11 reveal clear performance trends that validate the effectiveness of the proposed attention fusion strategy across different backbone depths and architectures.The ablation study shows a consistent improvement trend with increasing network depth, the advantage of ResNetV2 pre-activation over standard ResNet, and the limitation of VGG architectures without residual learning. Most importantly, replacing standard CBAM with the proposed LGF-CBAM yields the lowest false positive rate and the highest overall performance, confirming that the learnable gated fusion mechanism improves discriminative feature selection under complex field backgrounds.

3.3.1 Summary of all dataset performance on the proposed system.

The performance metrics of the Proposed System across five cocoa disease datasets demonstrate its strong generalization and reliability as shown in Fig 14 below. The highest performance was observed on the Cocoa_Pod_Disease_Gh dataset, with an accuracy of 98.95%, an impressive FPR of 0.0089, and a TPR of 0.9910, indicating excellent sensitivity and specificity. Its F1 score and PPV, both at 0.9911, affirm its balanced precision and recall. The Cocoa_Pod_Disease_Gh dataset also achieved the lowest loss value of 0.0232, indicating highly accurate predictions with minimal deviation from the expected outputs

Download:

• PNG
  larger image
• TIFF
  original image

Fig 14. a: Accuracy for All Datasets. b: Model FPR for all datasets. c: Model F1 Score for All Datasets. d: Model PPV Score for All Datasets. e:Model TPR Score, All Datasets. f: Model Loss Score for All Datasets.

https://doi.org/10.1371/journal.pone.0348147.g014

On the Cocoa Diseases (YOLOv4) dataset, the model also performed excellently, with an accuracy of 98.53%, slightly lower but still robust, and consistent F1 (0.9880) and PPV (0.9881) scores. The Cocoa Diseases (YOLOv4) dataset, recorded a loss of 0.0263, suggesting strong predictive performance. Similarly, performance on the Black and Borer Pod Rot dataset was strong, with 97.96% accuracy, showing the model’s ability to detect distinct disease types accurately, maintaining low error rates.

Although slightly lower, the Cacao Diseases in Davao dataset yielded 96.19% accuracy, still reflecting a reliable classification rate, despite a relatively higher FPR (0.0294). Moderate increases in loss were observed for the Cacao Diseases in Davao (0.0497) and Black and Borer Pod Rot (0.0479) datasets, reflecting slightly reduced model precision compared to the first two datasets, possibly due to differences in image quality, environmental factors, or dataset variability. The Coffee and Cocoa dataset showed the lowest accuracy at 94.00%, with a notable increase in FPR (0.0582), suggesting that this dataset may present more variability, noise, or challenging image features. Again, the Coffee and Cocoa dataset recorded a significantly higher loss value of 0.1711, indicating a greater margin of error in predictions. This higher loss could be attributed to potential noise in the dataset.

As shown in Table 12, the proposed model demonstrates very high in-domain performance across all five cocoa disease datasets, achieving accuracies between 94.00% and 98.95%, with consistently low false positive rates and strong F1 scores. These results confirm the model’s effectiveness when trained and tested within the same data distribution.

Download:

• PNG
  larger image
• TIFF
  original image

Table 12. Summary of all dataset performance on the proposed model.

https://doi.org/10.1371/journal.pone.0348147.t012

To rigorously evaluate cross-domain generalization, a Leave-One-Dataset-Out (LODO) strategy was adopted. For each experiment, the model was trained on four datasets and evaluated on the fifth, unseen dataset without fine-tuning, simulating real-world deployment where data distributions differ from the training environment.

The results in Table 13 show that although performance decreases slightly under domain shift, the model maintains strong generalization, with cross-domain accuracies ranging from 90.80% to 97.80% and an average accuracy of 95.26%. The mean accuracy drop across datasets is only 1.87 percentage points computed as the absolute difference between in-domain and cross-domain accuracies, indicating that the model retains most of its predictive capability even when exposed to unseen disease patterns and image characteristics.

Download:

• PNG
  larger image
• TIFF
  original image

Table 13. Cross-dataset performance.

https://doi.org/10.1371/journal.pone.0348147.t013

The largest degradation is observed on the Coffee and Cocoa dataset with 3.20% drop, likely due to the presence of coffee leaf diseases not represented in other datasets. Nevertheless, false positive rates remain relatively low across all domains, confirming preserved specificity. This robust cross-dataset performance can be attributed to the LGF-CBAM attention mechanism, extensive data augmentation, batch normalization, and optimal hyperparameter selection, all of which contribute to learning generalized disease features rather than dataset-specific patterns.

3.4 Comparison with existing related systems on cocoa pod disease identification

Table 13 benchmarks the proposed LGF-CBAM model against prior studies across multiple datasets. The model achieved 98.95% accuracy and an F1-score of 99.11% on the Cocoa Disease GH dataset.On the Cocoa Diseases (YOLOv4) dataset, LGF-CBAM recorded 98.53% accuracy and 98.80% F1-score, compared to 96% accuracy and 93.30% F1-score reported in [51] using EfficientNet-B0 and ResNet50. For the Cacao Diseases in Davao dataset, the proposed model achieved 96.19% accuracy and 97.06% F1-score, while [88] reported 91.79% accuracy and 82.08% F1-score using multiple CNN architectures. Additional class-wise comparisons show that [88] reported a TPR of 96.69% and F1-score of 82.08% on the Davao dataset, whereas LGF-CBAM achieved 97.06% for both metrics. [50] reported 91% across accuracy, TPR, FPR, and F1-score, and [46] reported an F1-score of 85.88%, all lower than the values achieved by the proposed model.

Earlier deep learning models such as EfficientNet-B0, ResNet50, VGG variants, MobileNet variants, and ResNet18 reported accuracies ranging from 83% to 96%. Where additional metrics were reported, their F1-scores indicate a weaker balance between precision and recall compared to the proposed approach. Traditional machine learning methods, including SVM, KNN, Gabor kernel convolution, and k-means clustering, generally recorded lower performance metrics, reflecting limited capability for complex visual disease recognition. By incorporating F1-score and related metrics alongside accuracy, Table 14 provides a clearer evaluation of classification effectiveness, showing that the LGF-CBAM model delivers more consistent performance across the evaluated related systems.

Download:

• PNG
  larger image
• TIFF
  original image

Table 14. Comparison of the proposed system with existing related systems.

https://doi.org/10.1371/journal.pone.0348147.t014

The metric used by most existing system was the accuracy of the model so that was the main metric used for the comparison. The accuracy of existing systems was quite high, so the proposed model had to achieve superior results. Methods used on the same dataset are used for the comparative analysis. Six systems were compared to the Proposed System on the Cocoa Diseases (YOLOv4) datasets. The proposed system achieved a remarkable accuracy of 98.95%, outperforming several recent models documented in existing literature as shown in Fig 15.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 15. Comparison of cocoa diseases (YOLOv4) dataset.

https://doi.org/10.1371/journal.pone.0348147.g015

When compared to [51], whose system recorded an accuracy of 96%, the Proposed System demonstrated a 2.95% improvement, reflecting its superior predictive performanceSuch a margin is large in scenarios that require machine learning, particularly those cases where precision is a key priority. The performance in [73] and [45] were also similar and were at 84.62% and 84.75, respectively. The system suggested exceeds these marks by more than 14 percentage points, which points out the significant progress in modeling design, training process, or data preprocessing. [50] had accuracy of 91 percent, and, [46] had 86.04 percent. Once again the proposed system has a strong lead with 7.95 percent and 12.91 percent respectively. One possible source of these improvements can be an improved model architecture or better informative features. Moreover, [50] had already obtained a 94 percent accuracy. Though this particular performance is deemed to be quite strong, the offered system demonstrates an improvement of 4.95% versus it, which is an indicator that significant work has been conducted since then in terms of algorithm development technicalities and performance optimizations.

According to the comparative performance of the Cacao Disease in Davao dataset, the proposed system had a much better result than the current approach developed in [88]. The Classification accuracy of the proposed system was recorded at 98.95% which is attributed to an effective and highly precise model that can accurately predict the condition of cacao diseases with few errors as indicated in Fig 16.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 16. Comparison on cacao disease in davao dataset.

https://doi.org/10.1371/journal.pone.0348147.g016

Comparatively, the model in [88] achieved the accuracy of 91.79% that is not bad but is well below the performance of the proposed system. More than 7 percentage points of accuracy difference show the increased capacity of the proposed model which may be affected by enhanced architecture, new training strategies, better regularization, incorporation of LGF-CBAM attention mechanism and optimization of new feature extraction techniques

The proposed system also demonstrates a major improvement over all other models when it comes to accuracy in the utilization of all the existing private datasets where the Proposed System achieved a high rate of 98.95%. In comparison, [42] recorded 89. 2% as the second-best accuracy rate followed by [89] (86.2%), and [47] (86.04%). Though relatively competitive, these models also underperform by over 10 points relative to the proposed system, which also is better at predicting the outcome based on the Fig 17.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 17. Comparison on private datasets.

https://doi.org/10.1371/journal.pone.0348147.g017

Other prior works such as [90] (84.44%), [92] (83%), and [49] (81%) demonstrate moderate performance, yet they lag considerably behind the proposed model. The study by [91] shows even lower accuracy (79.68%), and the earlier model by [43] achieved only 70%, suggesting that earlier approaches struggled with generalizability or handling complex image features.

4.1 Interpretation of key findings

The experimental results demonstrate that the proposed LGF-CBAM architecture achieves state-of-the-art performance across five diverse cocoa disease datasets, with in-domain accuracies ranging from 94.00% to 98.95% and F1-scores between 94.14% and 99.11%. Three key findings emerge from these results. First, ResNetV2-101 outperformed other pretrained base models, achieving 87.95% accuracy compared to VGG-16 (80.75%) and ResNet-50 (82.15%), with statistical significance confirmed via paired t-tests (p < 0.01). This confirms that architectural innovations, such as pre-activation residual connections and improved gradient flow, enhance feature learning capacity. Despite VGG-16 having a similar parameter count, its sequential convolutional design produces less discriminative features than the deeper residual architecture.

Second, the integration of the Learnable Gated Fusion CBAM (LGF-CBAM) attention mechanism substantially improved performance. On the Cocoa Disease GH dataset, accuracy increased from 87.95% to 98.95%, highlighting the role of adaptive attention mechanisms in emphasizing disease-relevant features while suppressing background noise.

Third, cross-domain (LODO) validation demonstrated strong generalization. The model experienced an average accuracy drop of only 1.87 percentage points under domain shift, maintaining above 90% accuracy even in the most challenging cross-dataset scenarios. This robustness is attributed to ImageNet pretraining, extensive data augmentation, and the attention mechanism’s ability to focus on disease-centric rather than dataset-specific features.

4.2 Comparison with existing literature

The LGF-CBAM model not only improves accuracy but also provides a balanced evaluation across multiple metrics. On the Cocoa Diseases (YOLOv4) dataset, the model achieved 98.53% accuracy and 98.80% F1-score, outperforming [51] (96% accuracy, 93.30% F1-score) and demonstrating better balance between precision and recall. On the Cacao Diseases in Davao dataset, it achieved 96.19% accuracy and 97.06% F1-score, surpassing ensemble-based CNN approaches by 14.98 percentage points in F1-score while maintaining lower inference complexity. Compared with traditional machine learning methods such as k-means clustering and SVM/KNN variants, the proposed system achieved higher accuracy (98.95%) and lower false positive rates (0.89%), emphasizing the superiority of hierarchical deep feature learning for complex visual disease recognition. By reporting accuracy, F1-score, TPR, and FPR, this study provides a more comprehensive benchmarking standard than many prior works.

4.3 Limitations and threats to validity

Despite strong performance, several limitations must be acknowledged. Dataset sampling bias may have influenced results, as The Cocoa Disease GH dataset was collected under real field conditions in specific locations in Ghana. The lower performance on the Coffee and Cocoa dataset confirms that greater visual variability and unseen disease patterns can affect predictions. Inter-disease similarity also posed challenges; most misclassifications occurred between Phytophthora and Moniliophthora rather than between healthy and diseased pods, indicating difficulties in fine-grained discrimination during early infection stages. Environmental variability, including lighting, occlusion, and camera angles, was not systematically varied, so extreme field conditions were not explicitly tested. Finally the computational demands remain high, with ResNetV2-101 requiring 1,337 minutes of training and 44.7 million parameters, limiting deployment on low-resource devices without compression.

4.4 Practical implications for agricultural deployment

The results have strong real-world implications. The model’s high true positive rates of 94% to 99% enable early disease detection, supporting timely interventions before significant crop damage occurs. Very low false positive rates, as low as 0.89%, reduce unnecessary fungicide application and associated costs. Minimal performance degradation under cross-domain testing indicates that a single trained model can be deployed across regions. The LGF-CBAM could be integrated into mobile-based agricultural extension services, allowing farmers and agricultural officers in remote areas to access accurate cocoa disease detection tools.

To further assess deployment feasibility, the computational profile of the model was examined with respect to inference cost and memory requirements. Although ResNetV2-101 contains 44.7 million parameters, inference per image was observed to be fast and stable, making real-time prediction achievable on modern smartphones and edge devices with moderate processing capability. The use of Global Average Pooling, an optimal batch size, and a compact 2048-dimensional feature representation significantly reduce memory overhead during inference. In addition, the architecture is compatible with model compression techniques such as pruning, quantization, and conversion to TensorFlow Lite or PyTorch Mobile formats, which can further reduce model size without substantial loss of accuracy.

Collectively, these findings suggest that the proposed system is not only robust in controlled experimental settings but also computationally practical for real-world agricultural monitoring, mobile deployment, and field-level intervention strategies in resource-constrained environments.

4.6 Conclusion

The proposed LGF-CBAM integrated with a ResNetV2-101 backbone addresses several persistent limitations in agricultural deep learning literature. While prior attention-based CNNs and transformer models often demonstrate high performance on controlled datasets such as PlantVillage, their generalization under real-field variability remains limited. Through multi-dataset evaluation and Leave-One-Dataset-Out (LODO) validation, LGF-CBAM demonstrates strong cross-domain robustness, with only a 1.87% average accuracy decline under domain shift, indicating improved resilience to variations in lighting, background clutter, and geographic diversity.

Unlike existing CBAM adaptations that emphasize either lightweight efficiency or domain-specific architectural fusion, LGF-CBAM introduces hierarchical local global feature interaction tailored specifically to agricultural imagery. By fusing fine-grained lesion cues with broader pod-level structural context, the model overcomes the common limitation of purely local attention refinement and better captures the multi-scale nature of plant disease manifestation. In addition to accuracy gains of up to 98.95%, the model maintains moderate computational demands relative to transformer-based or multi-branch attention systems. While ResNetV2-101 contains 44.7 million parameters, inference efficiency and compatibility with pruning, quantization, and mobile deployment frameworks position the model as practically deployable in resource-constrained agricultural environments. This directly responds to scalability and edge-deployment concerns raised in recent literature.

However, consistent with broader research gaps, the current framework remains limited to static image-based diagnosis, severity estimation, and early pre-symptomatic detection. Furthermore, while quantitative metrics demonstrate strong performance, qualitative metrics with deeper interpretability validation with plant pathologists and integration of agronomic expertise are necessary to enhance trustworthiness and ensure responsible field deployment. Overall, LGF-CBAM advances CBAM from a generic attention enhancement module toward an agriculture-aware, multi-scale feature refinement mechanism. By balancing robustness, computational practicality, and field adaptability, the proposed system represents a meaningful step toward reliable, real-world cocoa disease diagnostic tools. It supports early intervention, reduces chemical misuse, and improves crop management outcomes, ultimately enhancing farmer livelihoods and food security in cocoa-producing regions.

4.7 Recommendations

Based on the outcomes of this study, the following recommendations are proposed:

1. Future agricultural deep learning systems should incorporate adaptive attention mechanisms, such as LGF-CBAM, to improve sensitivity to subtle and complex visual features commonly found in plant diseases.
2. Stakeholders including governments, research institutions, and industry partners should invest in the development of diverse, well-labeled agricultural image datasets across multiple ecological zones to enhance model generalization and robustness.
3. Given its strong baseline performance, ResNetV2-101 is recommended for tasks that require an optimal balance between precision and recall.
4. Close collaboration with plant pathologists and agronomists during the data annotation and evaluation phases is essential to ensure biological accuracy and the practical relevance of AI-based diagnostic systems.
5. Efforts should be made to deploy trained models on mobile and edge devices, enabling real-time field diagnostics that are easily accessible to farmers and agricultural extension officers.

4.8 Future work

The following directions are recommended for future research:

1. Implement active or continual learning frameworks that allow the model to evolve by incorporating new field data over time, thereby improving adaptability and long-term relevance.
2. Design and deploy a user-friendly mobile application that integrates the trained LGF-CBAM model, with offline functionality, to facilitate on-field cocoa disease diagnosis in remote areas.
3. Develop temporal deep learning models using time-series imagery to enable early detection of infections and monitor disease progression over time.
4. Evaluate the model’s generalization capabilities across different geographical regions and apply transfer learning techniques to adapt the system for other economically important crops such as cassava, maize, and coffee.