A sequential attention-enhanced deep learning framework for robust potato leaf disease diagnosis under real field conditions

Abstract

Diagnosing potato leaf diseases from images collected in real-life field settings is challenging, mainly because of uneven lighting, complex backgrounds, and disease symptoms that are often subtle or visually inconsistent. In this study, a deep learning-based framework was developed to support potato leaf disease diagnosis, with particular attention given to improving generalization and interpretation. Several convolutional neural network (CNN) architectures were first examined under the same experimental conditions, and ResNeXt-50 showed the most stable overall performance. The model was then extended by applying efficient channel attention (ECA), followed by spatial attention adapted from the convolutional block attention module (CBAM). Test results indicate that this sequential attention design performs better than the baseline model as well as variants using only a single attention mechanism. Additional evaluation using 300 real-field images collected under different field conditions suggests improved robustness, while visualization results from gradient weighted class activation mapping (Grad-CAM) show clearer focus on lesion-related regions. Overall, the findings suggest that combining channel wise and spatial attention can improve both prediction reliability and interpretability, making the approach suitable for practical agricultural use.