Banana Leaves Nutrient Deficiency Detection using Latent Attention Convolutional Neural Network

Abstract

Early and accurate diagnosis of nutrient deficiencies in banana leaves is crucial for ensuring optimal crop health and maximizing yield. We present the Convolutional Latent Attention Network (CLAN), a unified deep learning framework that synergistically combines global context modeling and region-based localization for precise classification of nutrient disorders. CLAN begins with a lightweight hierarchical convolutional encoder that extracts multi-scale feature maps from high-resolution leaf images, culminating in a compact latent code. By applying self attention, the Latent Attention Refinement module accentuates key global features linked to deficiencies. The Region Proposal Network (RPN) operates on the highest-level encoder features to identify candidate areas of interest. These regions are pooled via ROIAlign across all encoder stages and processed through a MobileNetV2 backbone to generate detailed local descriptors. By concatenating the refined latent code with region-wise features, the CLAN classification head achieves robust identification of multiple deficiency classes. The proposed model achieves a 95.2% F1 score and processes images at 4.5 ms inference time, while remaining compact at only 1.97 million parameters. These results demonstrate that CLAN combines rapid convergence and strong generalization with real-time capability and minimal hardware demands, making it exceptionally suitable for deployment in field conditions and resource-constrained environments.