A Lightweight Self Attention Based Multi-Task Deep Learning Model for Industrial Solar Panel and Environmental Monitoring

Abstract

Environmental monitoring has become a serious topic of discussion and is gaining mass attention. The reason is the severe consequences of environmental depletion, which has led to circumstances like climate change, rise in floods and droughts, changed rainfall patterns, etc. So, various measures are being taken to protect the environment, like shifting to renewable and pollution-free energy alternatives, like solar energy, and handling the after-effects of disasters, like flood management and oil spill accident management. However, their identification still remains a huge challenge, which is laborious and extensive. Thus, this work proposed a lightweight and efficient segmentation model, SA U-Net++, for the automatic identification of solar panels and their associated defects, flood affected-areas and oil spill accident regions. The model's novel blend of level-wise self-attention modules is embedded with the revised bridge connections and the dropouts. It has helped in better efficient global context understanding and feature extraction from the inputs, besides maintaining the integrity of the training process and avoiding some major learning and run-time issues, like overfitting and memory exhaustion. Our detailed experiments demonstrate that the proposed model outperforms state-of-the-art models. The results confirm its high generalizability, cost-effectiveness, and robustness.