Mapping sub-pixel corn distribution using MODIS time-series data and a random forest regression model

摘要

Accurate and timely information of the extent and changes in corn cultivation are of great significance for agricultural production management, food security and global environment change studies. Due to the high temporal repeat interval, near global coverage and rich spectral bands, MODIS time series data has been demonstrated particularly suitable for detecting the seasonal dynamics of different crops. However, their inherently coarse spatial resolution limit the accuracies of corn identification in regions with small fields or complex agricultural landscapes. In this study, we investigate the potential of using the random forest regression (RF-r) model to map sub-pixel corn cultivation based on time-series of MODIS data. Corn in Heilongjiang province, China in year 2011 was selected as a case study. Five time series of vegetation indices (155 features) derived from different spectral channels of MOD09A1 data were used as candidate features for the RF-g model. The out-of-bag strategy and backward elimination approach were applied to select the optimal spectra-temporal feature subset for corn identification. These derived corn maps were assessed in two ways: (1) wall-to-wall pixel comparison with the Landsat-based reference map to evaluate the classification performance in terms of spatial distribution. (2) prefectural- and county-level comparison with census data to evaluate its classification performance in terms of area estimates. Results show 61 optimal spectro-temporal features for corn cultivation were selected, which achieved the highest classification accuracy, with squared R (R~2) of 0.7586 and the root mean squared error (RMSE) of 0.085. MODIS-derived corn cultivation area had good agreements with the census data, with R~2 of 0.73 and RMSE of 238.07 km~2 across 43 counties, R~2 of 0.83 and RMSE of 1155.57 km~2 across 12 prefectures. These promising results indicate the great potential of RF-g method in mapping sub-pixel crop distributions based on coarse spatial resolution images.