Modified U-Net based photovoltaic array extraction from complex scene in aerial infrared thermal imagery

Hotspot detection in infrared thermal (IRT) images obtained by unmanned aerial vehicles (UAVs) is essential for the reliability, safety and efficiency of photovoltaic (PV) plants. We observe that some objects in the complicated environment, which are outside the region of PV arrays, might significantly interfere with the performance of automatic hotspot detection. This work focuses on the problem of PV array extraction from complex scene, which has not been solved in previous literature. Existing studies have attempted image processing methods to extract PV arrays. However, these approaches rely on manually designed features and parameters, which are impractical for complex scene. In this paper, a deep learning model, namely modified U-Net, is proposed to solve the problem. The proposed model can learn features and parameters automatically. Base on the classical U-Net, several innovations are provided: (1) batch normalization layers are adopted to alleviate internal covariate shift problems; (2) “He initialization” is used to increase the robustness and speed up the convergence; and (3) “RMSprop” is adopted to update parameters adaptively. 1211 IRT images of a PV plant containing complex background are collected to verify the effectiveness of the proposed method. Extensive experiments are conducted to compare the modified U-Net with five deep learning models and three image processing methods. The results demonstrate that the proposed method performs better than existing methods. Cooperating with the modified U-Net for PV array extraction from complex scene, even a simple hotspot detection method can achieve the accuracy of 99.79% and the F1 score of 0.9548.