High-resolution Remote Sensing Image Classification Based On Deep Feature Learning

Remote sensing images(RSIs)are usually to record electromagnetic wave radiation information on the earth’s surface detected by sensors mounted on platforms far away from the ground,and are imaged from spatial and spectral dimensions.According to the different spectral resolution during imaging,RSIs are divided into multispectral,hyperspectral and ultraspectral.Different types of RSIs have been extensively used in land use analysis,natural environment protection,military security,etc.The classification of RSIs is fundamental in RSIs processing.Due to the characteristics of large amount of data,high resolution and the different sizes and shapes of surface objects,there are great challenges in the classification of RSIs.With the development of deep learning,a great quantity of deep learning-based methods was applied to the classification of RSIs.Different from the machine learning-based methods,deep learning-based methods have become the main research method of RSIs classification,because they do not need to manually design feature extractors,and are easier to obtain abstract feature.This thesis utilizes deep learning-based methods to study the classification of different types of RSIs data.The main work is as follows:1)According to the different resolution of remote sensing images spectral band,RSIs are divided into multispectral images,hyperspectral images and ultraspectral images.The characteristics of different types of RSIs,some progresses in the RSIs classification and deep learning are reviewed.2)A novel classification method based on multi-level feature cascade is proposed for multispectral remote sensing images.This method obtains context information to make the classification results more accurate by using multiscale resolution.At the same time,different levels of backbone networks are used to reduce the model’s complexity,which can obtain faster segmentation speed at the expense of less accuracy.Experiments verify the effectiveness on the publicly available RSIs data sets,Vaihingen and Potsdam.3)A hyperspectral image classification based on three-dimensional(3-D)adaptive sampling and iterative shrinkage-threshold algorithm is proposed for the classification of hyperspectral remote sensing images.This method first performs 3-D adaptive sampling in the original image,and then uses the Iterative Shrinkage Thresholding Algorithm(ISTA)to reconstruct the image in an unsupervised manner,where the features are extracted at different levels in the reconstruction processing,and are fused into the classifier.Experiments verify the effectiveness on the publicly available hyperspectral image data sets,Indian Pines,University of Pavia and Salinas.