Detailed Crop Classification Using Integrated Spectral,Temporal,and Spatial Features Derived From Remotely Sensed Images

Agriculture is the foundation of the national economy.Obtaining accurate and timely crop spatial distribution information is very important to guide agricultural production,rational allocation of resources and solve the problem of food security.Agriculture has the characteristics of wide coverage,regional differences,seasonal variations,low economic benefit per unit area and so on.From the point of view of technology and economy,it is extremely difficult to obtain the annual information of crop planting by the method based on field investigation.However,with the rapid development of technology recently,remote sensing is the effective method to solve this problem.Compared with the traditional statistical methods,remote sensing has many advantages on obtaining crop information,including wide coverage,short revisit period,and low cost of data acquisition.When combined with geographic information system and global positioning system,remote sensing cannot only extract the crop planting area,but realize the precise positioning of crop spatial distribution.In this study,according to the data characteristics of multi-spectral time series images,three improved feature extraction algorithms are used to extract the spectral features,temporal features and spatial features,respectively.And a comprehensive classification method based on the spectral features,temporal features and spatial features is proposed for crop classification.The experimental results show that the proposed classification method can effectively improve the accuracy of crop classification.In addition,the proposed method is applied to crop classification in Northeast China and achieve a good result.The main work and conclusions of this paper are as follows:(1)An improved feature extraction algorithm based on manifold learning is presented to solve the problem of data redundancy of hyperspectral data.On the one hand,“graph growing” strategy is used to select different number of nearest neighbors for each pixel adaptively,which can solve the problem that the number of nearest neighbors are difficult to be determined.On the other hand,the objective function of the original algorithm is modified,and the similarities between non-nearest neighbors are enhanced.(2)Laplacian eigenmap(LE)algorithm based on dynamic time warping(DTW)is proposed to solve the high dimensional problem of multi-spectral remote sensing image data.LE-DTW cannot only use all available data in time series,but can perform dimensionality reduction for time series data with unequal length.As a result,feature bands with equal length including most useful information is formed for the subsequent application.Actually,most algorithms in feature extraction and classification cannot use the data with unequal length.(3)Minimum spanning tree(MST)based on dynamic time warping(DTW)is proposed to extract spatial feature for multi-spectral remote sensing image data.MST-DTW can also use all available data in time series.And the segmentation map can be obtained when the length of time series data is unequal.(4)This study presents a classification method combined spectral features,temporal features and spatial features,named STS classification method.This classification method uses all available multi-spectral time series remote sensing images.First,the pixel-wise classification map is achieved using spectral-temporal features.Second,the spatial normalization is performed for the pixel-wise classification map using the segmentation map acquired by the spatial feature.Third,the final thematic map is obtained using majority voting strategy.Experiments on three major agricultural regions in the United States show that the STS method achieves higher accuracies than the traditional classification method based on a single feature.(5)A complete acquisition scheme for crop spatial distribution information is designed for three main crops(maize,rice and soybean)in three provinces of northeast China(Heilongjiang province,Jilin province and Liaoning province),including the original time series data acquisition,data processing,feature extraction and classification.By comparing the data with the Statistical Yearbook published by the National Bureau of Statistics,the results show that the proposed scheme complete higher crop classification accuracies.The scheme only needs to collect some training samples manually,and can be executed without any extra manual intervention.Therefore,it provides a basis for the study of crop classification with high precision and large scale.