Deep Learning-Based Solar Panel Cleanliness Detection for Autonomous Cleaning Robots Using Adaptive Decision Thresholding

Abstract

Keeping solar panels clean is important for efficient energy production, particularly in regions with high levels of dust, sand, and air pollution. Manual observation and schedule-based cleaning are inefficient and expensive, and are not suited to automated cleaning systems. This paper presents a visual classification model based on deep learning that aids decision-making in an autonomous solar panel cleaning robot by differentiating clean and dirty panel surfaces. Unlike studies that pursued multiclass surface classification or dimensionally-reduced image representations, this study formulates the problem as binary classification, in line with real-world robotic cleaning. A convolutional neural network with a pre-trained ResNet-50 backbone was fine-tuned through transfer learning to learn discriminative features from real-world solar panel images captured under varying conditions. The evaluation emphasizes decision reliability, analyzing class probability distributions and applying threshold-based classification to control cleaning-triggering sensitivity, so that detecting dirt is prioritized over unnecessary cleaning. Experiments showed that the proposed approach achieved a favorable precision-recall trade-off for each class, with an accuracy of approximately 77% and a macro F1-score of 0.76 on an unseen test dataset. Adjusting the decision threshold improved dirty-panel recognition without retraining the model. The contribution of this work is a practically deployable vision module for autonomous solar-panel cleaning systems.