Attention-Enhanced GCN with SPP for UAV-Captured Plant Imagery

Global agricultural productivity faces severe threats from plant diseases, which demand the development of effective and scalable detection methods. Traditional methods like manual inspection combined with handcrafted feature extraction demand extensive labor while proving ineffective for large-scale applications. The combined progress in Unmanned Aerial Vehicle (UAV) technology and deep learning (DL) now supports automated disease detection, yet current techniques fail to represent intricate spatial relationships and dependencies within plant images, which restricts their performance. This research document introduces an innovative framework that combines Graph Convolutional Networks (GCNs) with attention mechanisms, Spatial Pyramid Pooling (SPP), and skip connections to tackle UAV-based plant disease classification challenges. This framework utilizes multiple Convolutional Neural Networks (CNNs) to extract robust features while integrating multi-scale features through SPP and refining them with GCNs that include attention mechanisms and skip connections. The results from 5-fold cross-validation demonstrate the framework’s remarkable validation accuracy of 97.61%, F1-score of 0. 9848, the precision of 0. 9857, and recall of 0. 9842. These results achieve substantial performance gains over EfficientNet-B0 ShuffleNetV2 ResNet-18 and a basic GCN without attention. Significantly, the GCN with attention surpasses the simple GCN by around 2.4% in validation accuracy. This research demonstrates that integrating UAV imagery with advanced DL models establishes scalable soybean disease detection, laying the groundwork for precision agriculture applications.