Hybrid DenseRefine U-Net: A Novel Deep Learning Based Framework for Solar Panel Segmentation

The adoption of renewable energy sources is essential as the globe confronts the escalating impacts of climate change. From this perspective, solar energy is a prevalent source of energy that minimizes dependence on fossil fuels. However, thorough monitoring and documenting of photovoltaic (PV) installations require additional contributions for efficient planning and policy development. Thus, this study presents a hybrid DenseRefine U-Net model that integrates DenseNet's efficient feature extraction with U-Net's encoder-decoder framework, enhanced by refinement operations inspired by RefineNet. The proposed approach allows for improved segmentation of solar panels in aerial and satellite imagery by addressing the challenges of occlusions and varying light conditions. A thorough performance comparison was conducted against several deep learning-based segmentation models, including U-Net, RefineNet, and DeepLabV3+. The findings indicate that the suggested model surpasses current methods using two comprehensive datasets obtained from Jiangsu Province, China. The proposed model achieved a Dice coefficient score of 96.48%, an Intersection over Union (IoU) of 93.22%, and an overall accuracy of 98.14% on the first dataset. Additionally, the second dataset attained a Dice score of 95.23%, an IoU of 91.05%, and an overall accuracy of 97.54%.