A novel comparison of image semantic segmentation techniques for detecting dust in photovoltaic panels using machine learning and deep learning

Abstract

The reduction in photovoltaic (PV) panel efficiency is a significant concern, especially for the photovoltaic power stations (PPS) near different soil types and a high wind presence. A relevant interest has emerged in developing systems capable of recognizing and evaluating the state of PV panels without human intervention. This work analyzes three different approaches to address this problem using semantic segmentation. The first approach employs unsupervised learning, while the second utilizes supervised learning, applying Machine Learning techniques such as K-means, Gaussian Mixture Models, Random Forest, and Light GBM, as well as more rudimentary options like histogram segmentation and color spaces. The final approach utilizes Deep Learning models, testing different versions of the U-net architecture primarily designed for image segmentation tasks. The results showcase the model's performance in terms of accuracy, processing and training time, F1 Score, and Intersection over Union. It was observed that supervised models with Machine Learning algorithms achieved a perfect balance between performance and speed. On the other hand, the Deep Learning approach proved more effective when the input was not standardized and the image format was poorly defined.