KishiNet: A Lightweight Hybrid CNN–Transformer Architecture with an Attention and Out-of-Distribution Rejection Mechanism for Explainable Multi-Crop Plant Disease Detection and Mobile Deployment

Plant diseases remain a major threat to agricultural
output and food security in Bangladesh, where rural smallholders
often have limited access to trained plant pathologists. Deep
learning offers an alternative diagnostic route, but most existing
models are built for a single crop, are too heavy for lowcost mobile hardware, or behave as opaque classifiers that give
a label without any visual justification. This paper presents
KishiNet, a hybrid architecture aimed at efficient, multi-crop
disease diagnosis on-device. The model couples a MobileNetV2
backbone for lightweight feature extraction with a Convolutional
Block Attention Module (CBAM) that refines spatial and channel
features, followed by a compact Vision Transformer (ViT) head
that models long-range dependencies between leaf lesions that are
spread across a frame. The network was trained and evaluated
on a curated set of 17,990 images spanning 17 classes across
five economically important crops in Bangladesh plus a background/noise class, combining public repositories with roughly
1,000 field-collected samples. After MD5-based deduplication and
class balancing, training used a stratified 70/15/15 split under a
two-phase schedule. To reduce false positives caused by nonbotanical objects appearing in unconstrained field photos, an
Out-of-Distribution (OOD) rejection layer based on softmaxthreshold gating was added to the decision pipeline. On the
held-out test set, KishiNet reached 97.49% accuracy with a
weighted F1-score of 97.48%, and a macro-averaged validation
accuracy of 99.24% was obtained during model selection. GradCAM visualizations are used to expose the leaf regions driving
each prediction, and the trained network was exported to a 9.72
MB TensorFlow Lite binary that runs inside the KrishiBondhu
Android application, allowing offline inference in areas with
limited connectivity.