Dryland Crop Recognition Based on Multi-temporal Polarization SAR Data

Abstract

Dryland crops in China have a large planting area and wide spatial distribution, which contributes a lot to grain production. Accurate and timely acquisition of information on dryland crop types, planting areas and spatial distribution in dryland can provide an important basis for agricultural production and management, as well as the formulation of national food policy and economic plan, therefore, it is of great significance for promoting structural reform of agricultural supply side and national food security. The key stage of crop growth in the north of China is affected by cloud and rain weather, which makes it difficult to obtain sufficient effective optical data, and the current recognition accuracy of dryland crops based on polarized SAR data is low. In order to solve these problems, this study selected Jizhou City, Hebei Province as the research area, using the full-polarization RADARSAT -2 data of July 17, August 7 and September 24, 2018, and then selected three polarization decomposition methods (Cloude-Pottier decomposition, Freeman decomposition and Yamaguchi decomposition) and two classification methods (maximum likelihood and random forest) to construct 18 classification combinations. The identification of corn and cotton in study area was studied by using various schemes. Finally, the accuracy of dry land crop identification under various combination schemes was compared quantitatively with the ground survey data. The results showed that, Yamaguchi decomposition combined with maximum likelihood classification method was used on August 7 (flowering and boll period of cotton), and the classification accuracy was the highest (production accuracy was 78.98%). For corn, Yamaguchi decomposition combined with random forest classification method was used on September 24 (mature period of corn), and the classification accuracy was the highest (production accuracy was over 90%). I For the decomposition method, Yamaguchi decomposition has the highest classification accuracy among the three decomposition methods, followed by Freeman decomposition, Cloude-Pottier decomposition has the lowest classification accuracy; as far as the classification method is concerned, the maximum likelihood classification method has the highest recognition accuracy for cotton, but the random forest classification has the highest recognition accuracy for corn; in terms of the best identification phase, the flowering and boll period is the best recognition period for cotton, and the maturity period is the best recognition time for corn.. This study will help to improve the recognition accuracy of corn and cotton in fully polarized SAR data, and provide reference for the identification of multi-temporal dryland crops under complex planting structures.