Desert Remote Sensing Image Recognition Based On Multi-resolution Feature Fusion

Desert recognition is the fundamental for preventing and/or controlling desertification.Topographical features of desert images change constantly due to the uncertainty of desert terrain,illumination,and other properties.Desert images are usually characterized by large image size,large scale change and irregular location distribution of surface objects.Traditional methods have the problem of inaccurate recognition when processing complex desert images.In recent years,deep learning image processing methods provides a good solution for desert image recognition.However,the existing deep learning methods directly applied to desert samples still have problems such as insufficient utilization of multi-scale information and rough recognition boundary.To solve the above problems,this paper proposed a multi-scale residual network based on an attention mechanism(MSRCNet)to classify desert images.In the improved multi-scale residual module,convolution kernels of different sizes were used to extract multi-scale features of desert images,skip.connection was used to improve the utilization of multi-scale information and possible gradient disappearance.The attention mechanism module was proposed to enable the model to select the appropriate feature channel adaptively,and enhances the model's selection of target scale information.However classification methods used sliding windows usually had problems such as time.consuming calculation and inaccurate edge segmentation results etc.,Image segmentation methods could better avoid the above problems.This paper also proposed a Multi-resolution supervision network with adaptive weighted loss(Mrs Seg-AWL)to segment desert images.Mrs Seg first used a lightweight backbone to extract multi-scale features,then adopted Multi-resolution fusion modules to fuse the local and global information,and finally a multi.level fusion decoder was used to get the desert segmentation result.In this method,each branch loss was treated as an independent optimization task,AWL was proposed to calculate and adjust the balancing parameters of each branch.By giving priority to the easy tasks the improved loss function could effectively improved the convergence speed and the desert segmentation result.The experimental results showed that two improved methods had more accurate desert classification results,and had important guiding significance for achieving accurate desertification monitoring.