Research On Photovoltaic Detection In Remote Sensing Image Based On Multi-feature Fusion

Climate change is a global issue that all mankind needs to work together to face.At present,our country is implementing the "carbon neutral,carbon peak" policy.The realization of carbon neutrality is a new development concept implemented by China,and the solar and photovoltaic power generation industry will play a crucial role in in achieving China’s carbon neutrality goal.Compared with other countries,our country has the characteristics of vast area and has great advantages in photovoltaic power generation.However,compared with China’s regional conditions and market demand,photovoltaic power generation still has a lot of room for development.Photovoltaic spatial distribution is the main form of green energy development,while remote sensing monitoring is the only technical means to obtain large-scale photovoltaic spatial and temporal distribution information data.Photovoltaic information plays a fundamental role in many fields.Photovoltaic information plays a fundamental function in many fields,and photovoltaic information can also be applied to emergency response,smart cities,etc.It has an important role in many development fields such as sustainable urban development and urban planning.Remote sensing images from satellites have a regular update mechanism,and with regular inspections,changes in photovoltaic distribution can be obtained in a timely manner.It is therefore very important to practically detect,locate and identify photovoltaic stations and to obtain information on photovoltaic.Remote sensing satellites are becoming more widely used as science and technology advance.We can obtain more high-resolution satellite images,and depth learning detection algorithms are also increasingly used in remote sensing image processing.Many traditional remote sensing image recognition algorithms have been proposed and many advances have been made after careful research by local and foreign scientists.However,because the remote sensing photovoltaic image has the characteristics of irregular size of photovoltaic features,dense distribution of small targets,and large differences in environmental noise in different regions,the results of traditional recognition algorithms ion remotely sensed photovoltaic images are not satisfactory,and there is a problem that the detection and recognition of photovoltaic power plant targets is not very accurate.In order to solve the problem of slow as well as inaccurate photovoltaic mark detection,considering the characteristics of high resolution in remote sensing images and high noise of photovoltaic images,this thesis uses the method of deep learning to design photovoltaic target detection model and photovoltaic semantic segmentation model,and on this foundation,a remote sensing image fusion method was designed.Specific work contents are as follows:1.This thesis proposed an improved model of Yolov5s:Yolov5s_gc..In view of the characteristics of irregular size and dense distribution of small targets of photovoltaic features,using Ghostconv to replace Conv in the model’s backbone network Backbone,a new GhostC3 network was designed to replace the original C3 network module in the C3 network structure,improving detection efficiency and feature extraction ability of the model.The CA block attention mechanism module is embedded to improve the extraction of photovoltaic target featuresIn view of the high resolution of photovoltaic data sets,the loss function is changed from GIoU_Loss is changed to consider more comprehensive SIoU_Loss to improve the accuracy of model detection.Use the Arcmap map tool to manually tag the remote sensing photovoltaic images from different regions of the country and to select remote sensing photovoltaic images from regions with different environmental characteristics as detection data.From the experimental results it can be seen that,compared with Yolov5s original method,the improved algorithm mAP@0.5,accuracy P and recall R reached 97.5%,98.9%and 94.9%respectively,increased by 1.8%,1.7%and 5.8%respectively.The improved model can have better detection and recognition effects on remote sensing photovoltaic targets,and can further improve the detection accuracy and generalisability of the network with the primitive method,It is verified that the algorithm has a good effect on photovoltaic detection.2.This thesis proposed an improved model U-Net_rd of U-Net model for photovoltaic segmentation method.First,the encoder uses the Resnet50 residual network structure to replace the original U-Net encoder structure improves the degradation caused by the increase of model depth;Then,DANet dual attention mechanism is added to enhance the image feature learning capability between network channels;Finally,the ReLU function of the model is modified to Mish function for training to improve the accuracy of model training.The Labelme tool is used to manually label the remote sensing photovoltaic images of domestic regions,and select the remote sensing photovoltaic images of different regions with different environmental characteristics as the detection data.After experimenting with the results obtained,the new model can achieve 98.35%,98.13%and 96.6%respectively in the accuracy of PA,mPA and mloU evaluation indicators,which have increased by 1%,0.9%and 2%respectively.The improved model can have relatively good recognition effect on remote sensing photovoltaic targets,and can further improve the detection accuracy and adaptability of the network to fresh samples under the original method.It is verified that the algorithm has a good effect on photovoltaic segmentation and recognition.3.This thesis proposed an combined method with the image fusion method and deep learning.The ENVI map tool is used to preprocess the high-resolution panchromatic graphics and low-resolution multispectral images collected by the Gaofen-2 satellite according to the longitude and latitude information characteristics for atmospheric correction and orthophoto correction,and the processed images are geometrically registered and the mask tool is used for GS fusion operation.An evaluation index is employed to assess and compare the fused image with the original,and it can be obtained with a high spatial resolution and multispectral information.Input the fused image into the improved model,and the accuracy,recall and mAP@0.5.The evaluation indicators reached 98.3%,93.9%and 95.5%,respectively;The accuracy of PA,mPA and mIoU in semantic segmentation reached 97.34%,97.36%and 94.82%,respectively.In the test image,the target of photovoltaic power plant with complex environment has promising results.