Hyperspectral With High Spatial Resolution Remote Sensing Imagery (H~2 Imagery) Classification Based On Deep Learning

Hyperspectral remote sensing has the characteristics of image spectral integration,and its precise spectral resolution can accurately represent the attribute information of ground objects.At present,the spaceborne hyperspectral observation platform has been widely used in large-scale macro monitoring,but it faces the problem of insufficient spatial resolution in small-scale monitoring.In recent years,with the rapid development of the micro UAV platform and lightweight hyperspectral imaging spectrum technology,the integrated UAV hyperspectral platform has made it possible to simultaneously obtain hyperspectral(nanometer-level)high spatial(centimetre-level)resolution images(H~2images)by reducing the observation height,realizing the simultaneous acquisition of rich spatial geometric information and spectral radiation information of ground objects.However,the details of ground objects are gradually revealed with the significant improvement of the spatial resolution,which makes the H~2 images show extremely high spectral variability and spatial heterogeneity.As a result,the phenomenon of the same class with different spectrums occurs in large numbers;the intraclass variance increases significantly,and the statistical separability of the spectral information is more complicated to distinguish,especially in the distinction of ground objects with similar attributes(the different types of crops,the different subclass of the same crop,etc.).Therefore,the research on the“observation-processing-application”systematic theory for the classification and interpretation of H~2 remote sensing objects and the effective integration of the advantages of hyperspectral and high spatial resolution remote sensing is the frontier scientific issues in the field of optical remote sensing.Combining the characteristics of H~2 images and application requirements,this paper focuses on four aspects of“the benchmark dataset construction-global spectral-spatial fusion classification-subclass classification-cross-regional classification”.The main research contents and innovations of this paper are summarized as follows:(1)Self-constructed WHU-Hi benchmark dataset for H~2 image classification.The WHU-Hi dataset,collected by a UAV-borne hyperspectral platform for smart agriculture application,includes four types of agricultural scenes with single regional and three cross-regional hyperfine crop classifications.Moreover,the WHU-Hi dataset has an extremely high pixel annotation ratio(over 80%)and more than 22 kinds of ground objects to verify the accuracy and generalization of the H~2 image classification algorithm.(2)A global spectral-spatial fusion classification framework is proposed to alleviate the problem of intra-class variance.This paper first introduces a conditional random field model for spatial patching mechanism to integrate the global spatial context information and utilize virtual sample augmentation to alleviate model overfitting.Moreover,a novel spectral patching mechanism is proposed to fusion global spectral-spatial information with a fully convolutional network and use the pre-train model with a natural image dataset to alleviate the limited training samples.(3)A spectral-spatial-scale attention network is proposed for subclass objects classification based on H~2 imagery.For the challenges of similar spectral information and with hyperfine objects and extremely high intra-class variance in H~2 image,the spectral attention,spatial attention,and scale attention modules adaptively weight the different channels,different spatial pixels,and different scale perceptions of the feature map,thereby relieving the problems of the intraclass spectral variability,spatial heterogeneity,and scale difference in hyperfine objects classification with H~2 imagery.Finally,the AAM loss function is introduced to maximize the classification boundaries in cosine space,which can increase the feature distinction between different classes,and therefore improve classification accuracy.(4)A cross-regional classification method based on pseudo-label learning with confidence estimation is proposed to alleviate spectral distribution shifts.The high confidence pseudo-label constrained class-to-class context information is proposed to lessen the intra-class variance significantly.The AAM loss function is introduced to increase the inter-class distance.A pseudo-label learning strategy of confidence estimation is proposed to reduce the misclassification problem caused by the spectral distribution shift.(5)An integrated prototype system for UAV-borne H~2 image classification based on the WHU-Hi dataset and proposed classification methods.The performances and algorithm applicability of the prototype system are comprehensively analyzed in the WHU-Hi dataset,which will provide a reference for the research and application of H~2image classification.In summary,this paper aims to integrate high spatial and high spectral resolution information to conduct theoretical system research of“observation-processing-application”for H~2 image classification and achieve complete and accurate attribute information extraction from ground objects.