A DEFORMABLE RESIDUAL DEEP CNN WITH SPATIAL PYRAMID POOLING AND ATTENTION MECHANISM FOR MAIZE PLANT DISEASE DETECTION

Abstract

Early and accurate identification of maize leaf diseases is essential for timely crop management and yield protection. However, conventional deep learning models based on fixed convolutional sampling may have limited capability to represent the irregular shapes, varying sizes, and multi-scale characteristics of disease symptoms, while irrelevant background regions can further affect feature learning. To address these challenges, this study aims to develop an adaptive deep learning framework for robust maize leaf disease classification. A Deformable Residual Deep Convolutional Neural Network (Deformable Res-DCNN) integrated with Spatial Pyramid Pooling (SPP) and an Attention Mechanism is proposed. Deformable convolution enables adaptive spatial sampling for capturing irregular disease structures, while residual connections facilitate effective deep feature propagation. SPP extracts multi-scale feature representations, and the attention mechanism emphasizes disease-relevant features while reducing the influence of irrelevant regions. The proposed model is evaluated using the maize leaf subset of the PlantVillage dataset, comprising Gray Leaf Spot, Northern Corn Leaf Blight, Southern Corn Leaf Blight, Maize Dwarf Mosaic Virus, and Healthy leaf classes. Experimental results demonstrate that the proposed model achieves 98.60% accuracy, 98.10% recall, and 97.25% F1-score, outperforming the considered baseline models. The findings demonstrate the effectiveness of the proposed framework for automated maize leaf disease classification and its potential application in precision agriculture.