Research On Remote Sensing Image Classification Based On Contextual Enhancement And Gated Fusion

Remote sensing image classification is the process of identifying and classifying each pixel in a remote sensing image and labelling it with the corresponding target class.This technique is widely used in land use mapping,urban planning and disaster emergency response,so accurate and rapid classification of features is of great research value and significance.With the advancement of earth observation technology,the resolution of remote sensing images is getting higher and the volume of data is increasing significantly,while the limited feature representation capability of traditional feature classification methods can no longer meet the accuracy requirements of real-life applications in the face of massive high-resolution images.In recent years,deep learning has made breakthroughs in the field of computer vision,which can learn robust feature representations from a large amount of data and become the mainstream of research on feature classification of high-resolution remote sensing images today.This paper proposes two feature classification algorithms for optical high-resolution remote sensing images based on the basic knowledge of theories related to semantic segmentation of deep learning,combined with the features of optical remote sensing images,and develops feature classification assistance software based on them.The research content of this paper is organized as follows:1)To address the problem that complex scenes of high-resolution remote sensing images tend to lead to fragmented prediction of large scale objects and loss of small scale objects,this paper proposes a context-enhanced textual representation enhancement network(CRENet)based on context enhancement for remote sensing image classification.The method defines a superpixel affinity loss based on inter-superpixel category relationships,which enhances consistent prediction of large scale objects by supervising the network to capture correct global contexts conforming to the category relationships during the training procedure.In addition,a local feature alignment enhancement module is proposed for the large number of small-scale objects in remote sensing imagery.The low-level feature layers are firstly enhanced by context contrast block consisting of multiple layers of dilated deformable convolutions with different ratios,and the deformable convolution is applied to align the shallow feature layers with the deep feature layers,and then the aligned feature layers are fused to enhance the small target detection capability of the network.The experimental results show that the method significantly improves the segmentation accuracy of objects at different scales in remote sensing images,and can effectively solve the problem of missing small-scale objects in classification.2)To address the problems of high computational complexity and memory consumption of existing deep learning methods,a plug-and-play light-weight gated fusion network(Aligned Gated Network AGNet)based on Res Net18 is proposed for remote sensing image classification.The method is based on an encoding-decoding architecture.Its lightweight feature alignment module learns the semantic differences between feature maps of different layers through deformable convolution,and uses high-level features to guide the alignment of low-level features to eliminate the semantic gap.Considering the limited feature extraction capability of the lightweight skeleton network,a gated feature selection module is designed to selectively fuse aligned feature maps from different layers through a gating mechanism to enhance valuable information and suppress useless noisy information.The experiment results demonstrate the effectiveness and lightweight of AGNet,allowing for fast and efficient classification of largescale,high-resolution remote sensing imagery in the presence of hardware constraints.3)To meet the application needs of high resolution remote sensing image classification,a high resolution remote sensing image classification aid software is developed.The software includes a statistical analysis toolbox and integrates the two classification algorithms proposed above.