The work arrives at a moment when the stakes for crop health could hardly be higher. Global food production will need to roughly double by 2050 to feed a growing population, and climate change is accelerating the spread of insect pests and the plant diseases they carry. Rising temperatures, shifting rainfall, and higher carbon dioxide concentrations are already eroding natural biological controls, increasing the risk of crop failure. Fruit crops sit squarely in the danger zone: watermelon ranks among the world’s top five fruits by volume, with more than 100 million tons produced annually, of which China alone grows over 73 million tons. Grapes underpin a wine industry that produces roughly 75 million tons each year, and the guava market is projected to grow from 20.46 billion dollars in 2025 to 37.40 billion dollars by 2034.

Earlier attempts at automated leaf disease detection have struggled to balance accuracy against practicality. A fuzzy-logic classifier built on RGB color data from 200 watermelon leaf samples managed just over 67 percent accuracy. A random forest pipeline reached only 73.38 percent across 38 disease categories, and its performance degraded as the number of classes grew. More recent deep learning approaches have pushed accuracy above 98 percent — a DenseNet121 and EfficientNetV2B0 hybrid hit 98.40 percent on watermelon diseases, and a VGG16 variant reached 98.4 percent on grape leaves — but these models carry tens of millions of parameters, making them poorly suited to the edge devices, IoT sensors, and low-power phones that dominate real farming environments. Many also operate as opaque black boxes, offering farmers no insight into what the network actually saw.

LeafXNet attacks both problems at once. The architecture takes RGB leaf images of 224 by 224 pixels, normalizes pixel values to the range from zero to one, and passes them through four convolutional blocks with 64, 64, 32, and 32 filters respectively, each followed by max pooling that progressively shrinks the feature maps. After the fourth block, the network holds a 12 by 12 grid of 32-channel features. Here the design departs from convention: the feature maps are flattened into a sequence of 144 tokens, each 32 values long, and fed into a multihead self-attention layer with two heads. This mechanism computes queries, keys, and values, applies the scaled dot-product softmax operation, and lets every spatial position weigh every other one, capturing long-range dependencies that a purely convolutional stack would miss.

The attention output is reshaped back into its 12 by 12 by 32 spatial layout, now enriched with context-aware importance scores that highlight disease-relevant regions while suppressing background clutter. A global average pooling layer collapses each feature map to a single scalar, a dense layer of 128 neurons learns high-level abstractions, and a five-neuron softmax head delivers class probabilities. The entire network — convolutional extraction, attention fusion, and classification — totals 79,717 parameters, roughly 55 times smaller than NASNetMobile, the lightest of the baselines it outperformed. Training used the Adam optimizer with a learning rate of 0.001, a batch size of 32, and up to 50 epochs, on a workstation with an NVIDIA RTX 4080 GPU.

The evaluation rested on three independently sourced datasets. The watermelon set came from the Regional Horticulture Research Station in Lebukhali, Patuakhali, Bangladesh, with 1,155 original images covering healthy leaves, mosaic virus, downy mildew caused by Pseudoperonospora cubensis, and anthracnose caused by Colletotrichum orbiculare; augmentation through flipping, scaling, luminance adjustment, and random zooming expanded it to 6,930 images. The grape dataset, assembled in Nashik, India, with annotation support from domain specialists at the Grape Research Centre, contained 2,726 images of healthy leaves plus downy mildew, powdery mildew, and bacterial leaf spot, growing to 5,016 images after augmentation. The guava set, collected from three locations in Bangladesh, spanned anthracnose, canker, dot, rust, and healthy classes; after augmentation and subsampling to correct a healthy-class imbalance, it reached 7,856 images. All datasets were split 80:10:10 into training, validation, and test sets.

The results were striking. On watermelon, LeafXNet recorded 100 percent accuracy, precision, recall, and F1-score on the test set, correctly classifying every one of 693 test images. On grapes it scored 99.80 percent across all metrics, with a single misclassified downy mildew instance among 503 test samples. Guava proved hardest at 97.33 percent accuracy, with the dot disease class dipping to a 95.95 percent F1-score. Crucially, a robustness analysis across five random seeds showed the performance was no fluke: mean accuracies of 99.74 percent for watermelon, 99.76 percent for grapes, and 97.59 percent for guava, with standard deviations of just 0.0032, 0.0029, and 0.0086 respectively. The baselines told a different story — Xception managed only 66.75 percent on guava, InceptionV3 57.37 percent, and NASNetMobile 71.55 percent, while even the strongest baseline, DenseNet121, reached 82.45 percent.

Transparency was treated as a first-class requirement rather than an afterthought. The team layered three explainable AI techniques over the trained network: class activation mapping, Grad-CAM, and Grad-CAM++. CAM multiplies class-specific weights from the classification layer by the final convolutional feature maps; Grad-CAM instead derives importance weights from the gradients of the class score with respect to each feature map; Grad-CAM++ refines this further by incorporating pixel-level gradient information and second-order derivatives, producing sharper, more reliable localization maps. In visualizations, Grad-CAM++ consistently zeroed in on diagnostically meaningful patterns — the concentric ring-like lesions of anthracnose, the mosaic-like mottling of viral infection, and the rust patches along mid and side veins — while suppressing spurious activations on healthy tissue.

Those qualitative impressions were put to a formal test. A plant pathology expert, Dr. Mohammad Idris Ali Howladar, Principal Scientific Officer at the Regional Horticulture Research Station, independently rated 300 explanation maps — 100 images per fruit, 25 per disease class — on three criteria: morphological relevance, localization precision, and feature correspondence, each on a five-point scale. Grad-CAM++ earned an overall mean of 4.44, against 3.86 for Grad-CAM and 3.13 for CAM. A Friedman test confirmed the differences were statistically significant for all three criteria, with chi-square values of 389.77, 391.78, and 378.53 and p-values below 0.001, and post hoc Wilcoxon signed-rank comparisons with Bonferroni correction showed every method pair differed significantly. The resulting hierarchy — Grad-CAM++ above Grad-CAM above CAM — held across every criterion.

The computational case is equally compelling. LeafXNet requires 1.05 gigafloating-point operations per inference, compared with 9.13 for Xception and 11.47 for InceptionV3, and it posted the lowest latency on every dataset, reaching 5.85 milliseconds on grapes and 9.55 milliseconds on guava. The authors are candid about limits: the model distinguishes diseased from healthy leaves but does not separate diseases from visually similar physiological disorders, so it should serve as a decision-support tool rather than a standalone diagnostic system. Future work, they note, will extend the approach to more crops, integrate real-time monitoring, and test transferability to species the network has never seen. For now, the message resonates far beyond the lab: in agricultural AI, the smallest model in the room may be the one that finally earns a farmer’s trust.

Subject of Research: Lightweight explainable deep learning model LeafXNet for fruit leaf disease detection

Article Title: LeafXNet: A Lightweight Explainable Model With Integrated Attention Mechanism for Fruit Leaf Disease Detection

Article References: Al Masud, S. M. R., Eshita, F. S., Mahi, A. B. S., Mowla, T. J., Arnob, N. M. K., Saha, A. K., & Mridha, M. F. (2026). LeafXNet: A Lightweight Explainable Model With Integrated Attention Mechanism for Fruit Leaf Disease Detection. Plant Direct, 10(10), Article e70201. https://doi.org/10.1002/pld3.70201

Image Credits: AI Generated

DOI: 10.1002/pld3.70201

Keywords: LeafXNet, plant disease detection, explainable AI, Grad-CAM, convolutional neural network, precision agriculture, multihead attention, food security, Lightweight, Explainable, Model, Integrated