The Forest Land Detection Of Remote Sensing Image Based On End To End Deep Learning Method

Forest plays an important role in social and economic development.However,China’s forest is vast,so it’s difficult for common observation methods to grasp the internal situation of forest land in real time.Space remote sensing image has the characteristics of wide coverage,fast data acquisition and abundant image information,which is more suitable for forest land observation.Combining deep learning method with the technology of satellite remote sensing image can provide timely information for the investigation and planning of forest land resources by observation,so as to further enhance environmental monitoring and promote regional ecological construction.This paper proposes a set of remote sensing image forest land detection strategies based on end-to-end deep learning technology.The low detection accuracy caused by large differences,uneven distribution and different shapes in forest land categories can be solved from different angles by image segmentation method:1.Design the Most-sure fusion strategy.From the perspective of making full use of existing end-to-end deep learning models,the strategy compares the detection results of multiple existing models according to the pixel prediction probability value,and determines the final category.Through this strategy,the advantages of different models are retained to improve the accuracy of forest land detection of remote sensing images.2.Design a multi-branch regression layer fusion network.From the perspective of network optimization,use multi-scale outputs to carry out regression of multiple loss functions,and design a layer fusion up-sampling block to enhance the transmission of different types of characteristics,which can speed up network fitting,and improve the accuracy of forest land detection3.Optimize the forest land detection task of wide-field of view images,use the strategy of input segmentation to save memory,and optimize the "edge joining" phenomenon through mathematical morphology,so as to improve the efficiency and accuracy of end-to-end deep learning model in wide-field of view image.The experimental results show that the end-to-end deep learning model and strategy proposed in this paper improve the accuracy of forest land detection of remote sensing images.The Most-sure fusion strategy makes full use of the existing models.The multi-branch regression layer fusion network effectively improves the accuracy of forest land detection in remote sensing image.The methods of input segmentation and morphological optimization make forest land detection model easier to be carried out on different computers.The method proposed in this paper can effectively detect forest land in remote sensing images and has reference significance for ground cover detection in other remote sensing images.