Research On Semantic Segmentation Method Of Remote Sensing Image Based On Deep Learning

Semantic segmentation of remote sensing images is an important branch in the field of remote sensing image applications.The development of high-resolution satellite sensor technology has provided massive amounts of data and diverse data types for remote sensing images,leading to a wide range of applications for remote sensing image processing in areas such as fine agriculture,urban planning and military security.With the rapid rise of deep learning in the field of artificial intelligence,combining big data remote sensing images with deep learning has become a popular research area.In this research,we focus on high-resolution remote sensing images as the main research subject,study semantic segmentation algorithms based on deep learning,proposing two semantic segmentation models and testing and validating them on remote sensing datasets.The main work is as follows:For high-resolution remote sensing images,an attention-based multi-scale and multi-channel D-MMA Net model is proposed for segmentation of remote sensing image features.In order to solve the problems of blurred feature target boundaries and unclear edge segmentation of small objects in remote sensing images.A D-MA module with multi-scale feature extraction combined with adaptive channel attention mechanism is proposed to acquire multi-scale semantic information at different stages of the decoder and adaptively and efficiently fuse deep and shallow semantic information.In addition,the contour gradient learning CEM module is used to target the edge gradients of the image and as supplementary information,which helps in the prediction of irregular object edges and small sample object boundaries.The experimental results show that the proposed method has good segmentation results for small object boundaries.Based on the above research foundation of multi-scale feature extraction and the idea of Dense Net feature reuse,a Dense-Inception Net(D-INet)model for semantic segmentation of remote sensing images is proposed.The algorithm incorporates shallow information in the depth model and introduces an accompanying loss function to solve the problems of irregular object segmentation and misclassification of similar objects in remote sensing images.Firstly,a new Dens-Inception module is proposed by combining a multi-scale feature extraction module with the idea of feature reuse,which uses a wider dense connection to obtain more contextual information and meaningful features;secondly,a modified receptive field block(RFB)module is used as a shallow feature extraction method,using upsampling step for fusing features to enhance the representation of the feature extraction network;finally,the concomitant loss function used in the intermediate stage improves the transparency of the hidden layer of the neural network,and the jointly supervised neural network extracts a more recognisable feature map.The experimental results show that the network has good performance and can be effective for semantic segmentation of remote sensing images.