Optimal Segmentation And Deep Learning Based Classification For Remote Sensing Images With Spectral-spatial Information

Remote sensing images have recorded the information and changes of earth surface over the past decades.The abundant remote sensing data have been widely used in a variety of important areas,such as environment monitoring,urban planning,marine remote sensing,and mineral exploration.Efficiently and accurately analysing and understanding remote sensing images is the key to the further develop remote sensing technology.In the early days of the development of remote sensing image processing technology,researchers mainly processed remote sensing image data in a pixel-based manner.With the development of remote sensing technology,the spatial and spectral resolution of remote sensing images continue to improve.The pixel-based processing method cannot make full use of the rich spatial and spectral context information of remote sensing images,and the processing performance is not satisfied.To deal with this problem,researchers proposed spectral-spatial-based method.The spectral-spatialbased methods believe that the neighboring pixels are more likely to share a similar spectral information.Based on this idea,researchers developed a variety of spectral-spatial-based remote sensing image processing algorithms.In this dissertation,taking the spectral-spatial-based processing maner as the basic idea,we focus on remote sensing image segmentaion and classification tasks.By integrating the object-based analysis and deep learning-based techniques,we analysed the spectral-spatial HSIs segmentation and classification problems and developed corresponding algorithms.The main conttibutions of this dissertation are concluded as follows.(1)An object-based multi-scale segmentation optimization method is proposed for remote sensing images.Based on object-based image analysis,this method measures the local spectral heterogeneity of objects by calculating the spectral angle between inter and intra objects.Each object of the image is assigned a local spectral heterogeneity series,where the largest value of the series is considered as the optimal scale.Different from single scale optimization methods,the proposed method aims to directly search the optimal objects from results of all different scales and combine them into final segmentation results.The proposed method is able to adaptively select appropriate scales for different landscapes.Moreover,the proposed method can be applied in any other multiscale segmentation algorithms for further analysis.(2)A bi-level MRF segmentation method is proposed for remote sensing images.Traditional pixel based MRF model can lead to much segmentation mistakes.To overcome this problem,we introduced object-based analysis,and developed a pixel-based and superpixelbased MRF integrated segmentation method.In this method,a spectral histogram distance measure is proposed for estabilishing pairwise potential terms.The spectral histogram distance is used to measure the spectral difference between neighboring objects.In this model,the bilevel MRF model is applied to exploit the spectral-spatial information of remote sensing image for both pixel and object-level.The model is able to keep the boundary of objects and merge objects for optimizing segmentaion.(3)A convolutional deep extreme learning machine(ELM)is proposed for hyperspectal images classification.Due to the back propagation,traditional convolutional neural networks(CNN)usually need a long training time and consume much computational cost.The proposed method introduced ELM and designed a simple yet effective network.To exploit spectralspatial information of hyperspectal images,spectral and spatial domain CNN is used for feature learning.Furthermore,deep stacked ELM is used for spectral and spatial feature extraction.Different from traditional CNN,the proposed method does not calculate back propagation.The netowk is able to provide competitive classification performace and a fast training speed.(4)An efficient non-local deep learning framework is proposed for hyperspecteal images classification.Although CNN framework can extract spectral-spatial information by convolutional kernel,the extracted information is still limited in a fixed window and long-range information is ingored.Non-local based methods can overcome this problem but the computional cost is high.Therefore,an efficient non-local deep learning framework is proposed.The efficient non-local module computes the pixels’ relation in a criss-cross path.Using a recurrent operation,each pixel’s response is aggregated from all pixels of hyperspectal images.Moreover,the fully convolutional network(FCN)is used as backbone for hyperspectral image feature extraction.Experimental results demonstrate that the proposed framewok is able to extract local and non-local informatation and improve the classification performance.