Reliable crop type maps are vital for agricultural monitoring, ensuring food security, and environmental sustainability assessments. Coarse-resolution imagery such as MODIS are widely used for crop type mapping due to their short revisit cycles, which is advantageous for detecting the seasonal dynamics of different crop types. However, the inherently low spatial resolution may restrict their utility for mapping crop types in regions with heterogeneous agricultural landscapes. Agricultural statistics, which provide crop acreage information at different spatial and temporal scales, have the potential to improve crop type mapping from remote sensing. Yet, previous studies have often used agricultural statistics as reference data to evaluate the accuracy of satellite-derived crop type maps but have rarely utilized them to improve crop type distribution mapping. The utility of integrating agricultural statistics with satellite images to produce high-accuracy crop type maps is rarely explored. This study presents a methodology for mapping sub-pixel crop type distributions via the integration of MODIS time series and agricultural statistics. We tested our approach in Heilongjiang Province, which has the highest agricultural production in China. First, we used an optimized random forest regression (RF-r) model with training samples derived from high spatial resolution images (i.e. SPOT and Landsat) to predict the sub-pixel crop type distributions from MODIS time series. To optimize the RF-r model, an 8-day MODIS time series of five vegetation indices in 2011 were used as the candidate independent variables, and a backward feature elimination strategy was implemented to select the best variables for model prediction. Second, we developed an Iterative Area Gap Spatial Allocation (IAGSA) method to spatially reconcile the discrepancies between the crop acreage estimated from MODIS-based maps and the agricultural statistics. We found that the MODIS-derived crop fractions agreed with those derived from the high-resolution images, with R2 > 0.75 for all crop types, yet there was a clear discrepancy between the crop acreage estimated from MODIS and agricultural statistics. The sub-pixel crop type maps adjusted by IAGSA were not only consistent with the agricultural statistics for crop acreage, but also retained the spatial distribution patterns of the original MODIS-derived crop fraction. Our results suggest the advantages of integrating coarse-resolution images and agricultural statistics to map sub-pixel crop type distributions, and to provide consistent estimation of crop acreage. The presented methodology has the potential to map large-scale crop type extent across regions in a cost-effective way.

https://doi.org/10.1016/j.rse.2021.112365