A Lightweight Residual Separable Inception Network with Convolution-Only Attention for Explainable Solar Panel Soiling Classification

ABSTRACT This paper deals with the challenge of classifying solar panel images as clean or dusty, at a scale that matters to reduce photovoltaic-related costs and inspection efforts. To solve this challenge in a transparent and computationally efficient way, we propose a novel residual depth-wise separable lightweight Inception-based model created from scratch. The backbone is based on bottleneck (1×1) convolutions and spatially separable (1×3) and (3×1) convolutions grouped in multi-branch Inception style residual blocks, and is also informed by a convolution only attention module that merges bottleneck and spatially separable convolutions to produce feature re-weighting maps without utilizing pre-existing attention frameworks such as SE or CBAM. The model is learned on realistic solar panel dust data with disjoint training, validation, and testing sets. The images are rescaled at a fixed resolution and augmented through geometric transformations so as to enhance generalization capabilities. Quantitative experiments test the trained network on standard classification metrics, with the addition of confusion matrices and precision-recall curves to observe the performance of each class in detail, including accuracy, precision, recall, F1-score and ROC–AUC. Grad-CAM and LIME are embedded as XAI tools to show which areas of the solar panel images lead the model to its predictions. The results, in general, suggest that the lightweight and interpretable network design is capable of capturing discriminative dust related patterns, while the XAI analysis discloses the potential and existing limitations, as well as clear directions on the further enhancement of the methodology.