Impact of Data Quality on defect classification in electroluminescence images of Photovoltaic Cells

The influence of data quality on the efficiency of inferences is a fundamental issue in machine learning, and the automatic analysis of photovoltaic (PV) images is no exception. The quality of electroluminescent (EL) images of PV cells often suffers due to economic constraints in large-scale PV plant evaluations. Therefore, the limitations in data quality must be considered when assessing task performance, such as the classification of PV cell defects. This work investigates the effect of data quality on a classification problem by utilizing several large datasets of EL images. Our approach involves simulating degradation in image quality and ground truth, allowing us to measure the impact on classification performance. Our results quantify the decrease in classification performance caused by noise, out-of-focus images, varying resolutions, inaccurate module-to-cell image processing (slicing), and the size of the training dataset. Furthermore, we employ blind, repetitive ground truth labeling to evaluate the impact of human bias, specifically the rate of expert disagreement.