Deep Learning-based Image Segmentation Research On Photovoltaic Panels For Remote Sensing

Growing global energy demand has resulted in a severe drop in traditional energy sources,whereas solar photovoltaic(PV)power generation is less expensive and more profitable,contributing to the rapid rise of the PV industry globally.With the PV industry’s rapid development,it is critical to get installed PV capacity and location in order to optimize power system design.Satellite images have a broad range and generalization,which aids in understanding the dynamic changes of ground phenomena.Deep learning techniques have outstanding performance in tasks related to remote sensing images.Therefore,by leveraging the combination of deep learning techniques and high-resolution remote sensing images,it is possible to effectively extract PV panels.However,the deep learning model is still susceptible to erroneous detection or fuzzy edge segmentation of PV panels in the segmentation process,nevertheless,because of the small number of data samples and the complicated background of remote sensing images containing them.To explore the issues associated with the segmentation of photovoltaic panels,the following work has been completed in this paper:(1)In view of the unbalanced photovoltaic categories in the domestic photovoltaic datasets,particularly the poor spatial resolution remote sensing images.This paper builds a remote sensing picture dataset of scattered PV panels to provide the model greater generalization capabilities.First,publicly available distributed PV data with geographic coordinate information are gathered.Following that,the PV samples are cleaned,labeled and visualized.Furthermore,these datasets are data enhanced and segmented to serve as a foundation for future experiments.(2)This research uses segmentation networks in the form of encoders and decoders to optimize a conventional DeepLabv3+ network.Skip connections combine multi-layer features to integrate rich semantic information and aid in image recovery.To capture crucial spatial and channel information,global convolution and self-attentiveness methods are used concurrently.Finally,to recover the lost partial channel information,the channel fusion module is proposed.In the publicly available PV datasets containing PV01,PV03 and PV08,the optimized model achieves 87.02%,92.98% and88.43% IoU,respectively.The experimental results show that the model can accurately segment PV panels for high-resolution remote sensing images,and it can especially better safeguard centralized PV panels’ edges.(3)In extracting distributed PV panels from images with poor spatial resolution,the proposed multi-level feature fusion-based PV panel segmentation algorithm is capable of dealing with the problem of blurring edges.A self-convolution-based approach for extracting PV from high-resolution remote sensing images is provided.A U-shaped fine-grained capture module is built to allow the model to extract and merge more effective features.This paper designs the self-convolution to generate convolution kernel by using the information of the feature map itself and apply it to the convolution part in the U-shaped fine-grained information capture module,so that it can adapt to a wider range of data features and enhance the module’s feature expression ability and generalization ability.Furthermore,the designed resolution augmented perceptual information module extracts local information by increasing the feature map resolution and using small convolution to better obtain PV panel edge information and achieve accurate PV panel segmentation in low spatial resolution remote sensing images.Finally,experiments are carried out on three publicly accessible and distributed PV datasets,and the experimental results show that the model in this paper has a higher PV extraction ability than other networks,with average accuracy PA,CPA,IoU,Recall and F1 scores of 98.13%,93.67%,90.1%,92.09% and 92.87% respectively.