Deep Learning Based Semantic Segmentation Method Of Remote Sensing Image

Remote sensing earth observation technology is a fast and efficient method of data acquisition,and the obtained image is widely used in many industries,including forestry.Semantic segmentation is one of the main applications in forestry remote sensing image analysis.Compared with image classification,it can obtain more abundant semantic class information in image and provides important decision-making information for ecological monitoring,returning farmland to forest and other tasks.It is of great significance for forestry to study the fast and accurate semantic segmentation method.Due to the disordered distribution and irregular shape of trees in remote sensing images,U-Net,a classical semantic segmentation model based on deep learning,cannot achieve high segmentation precision for trees.In addition,U-Net also has problems such as high model complexity,time-consuming calculation and poor semantic segmentation precision of class-imbalanced dataset.For the above problems,we perform multi-step structural optimization on the classical semantic segmentation neural network U-Net and propose a neural network STS-Net(Simplified Tree Segmentation Network)for tree/background two-class semantic segmentation task.For the class imbalance problem existing in semantic segmentation dataset,we propose a class-sensitive weighted loss function.For the problem of poor smoothness of large-scale semantic segmentation map obtained by splicing remote sensing tile semantic segmentation results,we propose an overlap prediction post-processing algorithm.The main work in this paper are summarized as follows:(1)In order to improve the semantic segmentation precision,reduce the model complexity and speed up the calculations of the U-Net,we perform 4-step structural optimization to optimize the neural network and propose STS-Net semantic segmentation neural network.Improvements include(a)replacing the standard convolution layer with a compact convolution module,(b)removing unnecessary computation modules in the network,(c)using residual connection to enhance the convolution module,(d)using improved cascading atrous convolutions to enhance the convolution module sequence.The experiment results show that compared with the classical semantic segmentation network U-Net,STS-Net achieves higher semantic segmentation precision with less model weight and faster running speed.(2)For the class-imbalance problem in remote sensing image semantic segmentation dataset,we propose a class-sensitive weighted loss function.The proposed loss function uses the semantic segmentation precision of each pixel in the current time neural network as a variable weighting factor to improve the classical cross entropy loss function,which can reduce the proportion of pixels that have reached higher semantic segmentation precision in the average value of the loss function and therefore implement class equalization processing.Experiment shows that compared with the classical cross entropy loss function,the proposed class-sensitive weighted loss function can improve the segmentation precision of the neural network for small-sample pixels in the class-imbalanced dataset.(3)After using the remote sensing tile semantic segmentation result to be spliced into large-scale semantic segmentation map,the problem of local semantic segmentation discontinuity in the map appears.For this,we propose an overlap prediction post-processing algorithm.The algorithm uses a single-model multi-data copy processing strategy.For the overlapping regions of adjacent remote sensing tiles,the segmentation result of the neural network for each pixel copy of the same pixel in each tile are first accumulated,and final semantic class of the pixel in large-scale semantic segmentation map is then obtained.Experiment shows that the proposed algorithm can further improve the semantic segmentation precision of trained neural networks on test dataset.