Advances in crop fine classification based on Hyperspectral Remote Sensing

Abstract

Classification and recognition of crops is an important prerequisite for crop yield estimation and crop growth monitoring. Rapid and accurate acquisition of crop type, spatial distribution and area information can provide basic basis for crop planting structure optimization and structural reform of agricultural supply side. It is of great significance to the formulation of agricultural policy, the development of social economy and the guarantee of national food security. In recent years, hyperspectral remote sensing has been able to fine classify crop types and varieties and obtain spatial distribution maps and planting structure information of crops by virtue of its many bands, abundant spectral information and sensitivity to small spectral differences among ground objects. This paper summarizes the application of hyperspectral remote sensing in crop fine classification, summarizes the hyperspectral data sources commonly used in crop fine classification at home and abroad, such as Hyperion data, environmental satellite data, CASI data and OMIS data, and analyses the applicability of various data. Meanwhile, the methods of crop fine classification using hyperspectral remote sensing are summarized, including decision tree classification, support vector machine classification, multi-classifier integration, spatial-spectral feature classification, hyperspectral data and radar data fusion classification, and the characteristics of various classification methods are analyzed. It was found that the classification accuracy of crop fine classification based on hyperspectral data was higher (better than 90%). But there are still some shortcomings: (1) At present, scholars at home and abroad focus on areas with simple planting structure. Most of the crop types in these areas are rice, wheat and other large-scale food crops, but less on cash crops such as sesame, rape, peanut and so on. (2) Hyperspectral remote sensing has high classification accuracy for regions with fewer crop types, but the classification accuracy needs to be improved in regions with many crop types. (3) Hyperspectral data has a high dimension and a large amount of data processing workload, which is not suitable for fine classification of crops in large-scale areas. Future research directions: (1) Expanding the scope of hyperspectral remote sensing monitoring objects, mainly cash crops. (2) Selecting areas with complex planting structure, fragmented plots, fluctuating topography and various crop types for fine classification of crops. (3) Attaching importance to the essential features of hyperspectral remote sensing fine classification and finding a stable classifier which is generally suitable for crop fine classification. (4) The mechanism of crop fine classification using hyperspectral remote sensing and the method of multi-source data fusion need to be further studied.