Research On Forest Detection Algorithm Of Remote Sensing Image Based On Deep Learning

Forest land is widely distributed and rugged in China,so it is difficult to master the internal conditions of forest land in real time by ordinary observation methods.UAV remote sensing image has the characteristics of low acquisition cost,rich image and fast data acquisition,which is more suitable for forest land observation than other methods.Deep learning is an intelligent learning method.The combination of deep learning method and UAV remote sensing image technology can provide timely information for the investigation and planning of forest resources,so as to further strengthen environmental monitoring and promote regional ecological construction.The main contents of this paper are as follows1.In order to solve the problem that the existing methods lose the details of forest detection results seriously,this paper proposes a network structure based on multi-scale feature fusion.The network structure mainly includes the dense jump connection of feature fusion of low-level information and high-level information,and the empty space pyramid pooling module which can increase the receptive field of the network and extract features from multiple angles.Firstly,the encoder is used to extract the global information.Secondly,the decoder is used to restore the spatial resolution of the image.Finally,the classifier is used to segment forest land and non forest land.Compared with the classical segmentation network experiments,the results show that the multi-scale feature fusion method can solve the problem of detail loss.2.To solve the problem that the edge of woodland detection in UAV remote sensing image is not precise enough,this paper proposes a woodland detection model of RA-Unet based on residual attention mechanism,improves the coding decoding structure in U-net,and realizes end-to-end pixel level semantic segmentation.The model mainly includes resblock module,channel attention fusion module and multi-scale feature module.The resblock module is used to deepen the network layers and increase the nonlinear characteristics of the model;the channel attention fusion module fuses the context information to improve the attention to the edge;the multi-scale feature module strengthens the information extraction to prevent information loss.Experimental results show that RA-Unet algorithm has better segmentation accuracy than U-net algorithm,and all segmentation indexes are improved.Compared with the multi-scale feature fusion network proposed in this paper,RA-Unet has better effect on the edge.