A method for monitoring cropland retirement using Landsat images and time series subsequence of cropland probability was presented. First, random forest was used to classify the statistical values of intra-annual Landsat images to obtain the probability of belonging to cropland for each pixel. These cropland probabilities obtained constitute annual time series of cropland probability. Next, the time series of cropland retirement (cropland to non-cropland) and other landcover change were analyzed to obtain the time series subsequence representing cropland retirement, that is, the characteristic subsequence. Finally, the distance between the cropland probability time series of an unknown pixel and the characteristic subsequence of the cropland retirement was calculated to extract the area and timing of cropland retirement. The proposed method was validated in cropland retirement mapping using Landsat time series of Tumuertai Town, Chahar Right Back Banner, Inner Mongolia Autonomous Region. The results shows that compared with existing methods, the proposed method produced higher accuracy in the extraction of both the area and timing of cropland retirement. 

提出一种利用Landsat数据和时间序列子序列的退耕监测方法。首先利用随机森林方法, 对每年的Landsat数据统计值进行分类, 得到每个像元属于农田的概率, 由每年的农田概率构成年际的农田概率时间序列; 然后, 对退耕(农田变为非农田)及相关地物类别的农田概率时间序列进行分析, 得到代表退耕的时间序列片段, 即特征子序列; 最后, 计算未知像元的农田概率时间序列与退耕的特征子序列之间的距离, 提取退耕区域和退耕时间。利用内蒙古自治区察右后旗土牧尔台镇地区的多年Landsat时间序列影像验证该方法的有效性, 结果表明, 与现有的方法相比, 该方法在退耕区域和退耕时间提取方面均获得更高的精度。