Customized Graph-Cut Segmentation with Superpixels for Snowfall Detection on Solar Panels

Snow accumulation on solar panels reduces their performance in energy generation regions. Hence certain image processing methods were developed for segmenting snow from the normal areas of solar panels. Graph Cut and Flood Fill are the popular semi-automatic method used for this application where it relies on detecting an efficient region for marking the segmentation process. Similarly, the manual intervention fails to provide consistent results in such methods. Therefore, a customized graph cut methodology is proposed in this work by combining super pixel-based feature extraction and k-means clustering. At first, the input images are preprocessed by dividing the pixels to form super pixels as a group to reduce the computational complexity without losing the meaningful information. Features like intensity, edge density, texture, and color are extracted from these super pixel groups for segmenting the snow-covered regions on a solar panel. Further, the extracted features are normalized by a k-means clustering algorithm for identifying the highest intensity regions to observe the snow-covered regions. A graph cut-based segmentation technique is applied at the end to ensure the precise boundary delineation. The work is also incorporated with several post-processing techniques for improving the segmentation mask quality. This proposed method is evaluated against Graph cut and flood fill and satisfied with an average dice similarity coefficient of 93.09% and 95.11%.