Research On Multi-label Classification Of Optical Remote Sensing Image Based On Deep Learning

Recent decades have witnessed continuous development of Remote Sensing(RS)technology,such successes have led to significant growth of optical imaging satellites.Due to the need to precisely analyze massive RS resources,RS image classification has gained increasing attention.As one of the main methods of earth observation,the scientific management and efficient utilization of optical remote sensing image is of great importance.And multi-label classification for a large number of optical remote sensing images can automatically interpret image information and generate image annotation in batch,which is of great significance in many theoretical research and practical applications.In this thesis,the task of multi-label classification for optical remote sensing based on deep learning is studied.The main contribution and research results are as follows:1.To address the problem of complex geometric transformation and large label difference in optical remote sensing image,deformable convolution(DC)unit is introduced into our multi-label image classification task,which extracts image features of geometric invariance and adaptive receptive field,providing with excellent features representation for multi-label classifier.2.To take the problem of multi-label classification as the classification problem of graph nodes,a multi-label classification algorithm DCN-GNN(Multi-Label Classification with Deformable Convolution and Graph Neural Network)is proposed,which can model label correlation explicitly and predict all kinds of labels at the same time.The algorithm utilizes the attention mechanism of semantic guidance to decouple the image features,and then represents the semantic features corresponding to each label and the correlation between labels into a graph structure in a data-driven way.Then with GNN,the semantic information is propagated and updated between the adjacent labels,and a variety of semantic staggered image features are obtained,which are sent to the final multi-label classifier.DCN-GNN algorithm is highly interpretable,which effectively reduces the "semantic gap" between image features used for multi-label classification and human semantic understanding of images;3.To take the problem of multi-label classification as sequence prediction problem,a multi-label classification algorithm DCN-RL(Multi-Label Classification with Deformable Convolution and Reinforcement Learning)is proposed to implicitly model the correlation and sequential prediction of all kinds of labels.This algorithm uses a reinforcement learning method to make agents learn a strategy that outputs follow-up label according to image features and predicted labels.DCN-RL simulates the cognitive mechanism of human beings and generates labels for images in the order of easy first and difficult second.This sequential prediction method is able to capture the high-order correlation between labels and improve the performance of multi-label classification tasks;4.The results of contrast experiment and visualization on multiple datasets show that the deformable convolution operation has a high agreement with the optical remote sensing image,and the explicit and implicit label correlation modeling mechanism in DCN-GNN and DCN-RL can also improve the classification performance.The overall performance of the two algorithms is better than the existing methods of optical remote sensing multi-label classification.