| With the continuous development of remote sensing technology,more and more well-performed remote sensing platform data can be obtained.How to effectively utilize the new remote sensing image data to large-scale crop interpretation is a hot issue in the field of remote sensing research.The new space remote sensing platform has a better performance in crop classification tasks than the old one,because of the higher temporal,spatial and spectral resolution.However,access to optical remote sensing images is often susceptible to weather,especially in the vast southern regions of China,where it is cloudy and rainy and there is a serious lack of image data.Therefore,how to reduce the uncertainty in the process of crop remote sensing interpretation and improve the efficiency of data utilization in the absence of sufficient long-term sequence images is one of the key issues that need to be solved by agricultural remote sensing at present.This study addresses the above problems by using multi-temporal and hyperspectral remote sensing image data combined with machine learning models to investigate the crop classification problem under the lack of time-series image conditions in response to the cloudy rain in the Hubei study area.Specifically,the content and innovation work of this study is summarized as follows:(1)From the characteristics of different remote sensing data sources,we summarize the current research status and problems of existing crop classification,and propose to make comprehensive use of the optimal time window of remote sensing images,as well as mining the spatial-spectral information of key time phases for the task of crop classification,and study the effect and significance of these two strategies in classification from these two aspects.(2)The selection of optimal temporal window for crop classification was demonstrated,based on multi-temporal remote sensing images.Based on the real samples on the ground and the agricultural statistical data of the study area,this study investigates the optimal temporal window of crop classification by arranging and combining the images obtained in different periods,combined with the calculation of the importance of time-spectral features.(3)To argue for the substitutability of multitemporal images in the case of acquiring hyperspectral data in the absence of time-series data,and to propose a method for mining empty-spectral information using three-dimensional convolutional neural networks.In this study,a multi-scale three-dimensional convolution neural network method is proposed to effectively utilize the spatial-spectral information of hyperspectral images through multi-level extraction of spatial-spectral information.In addition,compared with multi-temporal images,the effectiveness of extracting spacespectral information is demonstrated from the experimental results.Aiming at crop remote sensing classification,this paper studies the application of the temporal-spatial-spectral features to large-scale crop classification,and reveals the performance effects of different features in classification tasks.Which has important scientific value and practical significance for the realization of large-scale crop mapping and management. |
