基于MODIS时间序列及物候特征的农作物分类

论文以2012年6月至2014年6月期间的MOD09Q1及2013年四五月的MOD09A1为数据源,合成归一化植被指数(NDVI)和归一化水体指数(NDWI),利用TIMESAT软件对NDVI时间序列数据应用分段高斯函数拟合方法重构NDVI时序曲线,并获取7个物候特征(Phenology,以下简称PH,包括生长季始期,生长季末期,生长季长度,NDVI振幅,NDVI左导数,NDVI右导数,生长季期间的NDVI积分).结合Landsat 8 OLI遥感影像,中国第二次土地调查数据和实地采样样本数据,根据2013年多种地物平滑后的NDVI曲线特征,将年NDVI最大值低于0.5的水体和建设用地掩膜去除.为了获取研究区农作物的最优分类方法,采用分层分类:首先对平滑后的46个NDVI时序数据进行支持向量机(SVM)分类,得到农用地等分类信息;其次利用平滑后的46个NDVI波段,7个物候参数及6期归一化水体指数相互组合,对农用地进行支持向量机分类提取3种农作物的分布信息.经不同波段组合分类对比可知,分类总体精度及Kappa系数的关系为:NDVI+NDWI>NDVI+PH+NDWI>PH+NDWI>NDVI+PH>NDVI>PH.研究结果表明,遥感数据波段的增加不一定带来较高的分类精度;论文中归一化水体指数有效地提高了水稻的分类精度.此外,辅以物候特征对农作物分类也具有一定的可行性.

Crop Identification Based on MODIS NDVI Time-series Data and Phenological Characteristics

Agriculture is the foundation of the national economy. Identification of agricultural information by using remote sensing technique in real-time have been a hot topic. This paper aims to study the distribution of the main crops (soybean, corn, rice) effectively in large scale. Firstly, with the Asymmetric Gaussians method of TIMESAT software, the MOD09Q1 datasets with 250 m resolution were used to filter and reconstruct the time-series NDVI curves. Then seven phenological characteristics (start time of the growth season, end time of the growth season, length of the season, amplitude of NDVI, left derivative of NDVI at the beginning of the growth season, right derivative of NDVI at the end of the growth season and integral of NDVI during the growth season) were extracted. Secondly, to analyze the characteristics of time-series NDVI curve of vegetables, water and construction land were masked off because their maximum NDVI values were less than 0.5. Then in order to get the optimal classification accuracy of the crop land, hierarchical classification method was conducted as below: 1) using SVM classification to extract agricultural area based on the time-series NDVI data; 2) using SVM classification to identify three crop classes (soybean, corn, rice) with different combination of three bands (NDVI: NDVI bands; PH: phonological bands; NDWI: NDWI bands) on the basis of the first step. We compare the Overall Accuracy and Kappa coefficient of different combinations, and the result was as below: NDVI+NDWI>NDVI+PH+NDWI>PH+NDWI>NDVI+PH>NDVI>PH, the combination of NDVI+NDWI being the best. It was found that higher dimensions won't bring higher accuracy necessarily, and the application of NDWI can improve the overall accuracy of rice effectively. In addition, it is workable to identify the crop types with the help of phonological information.