LAUSR

ABSTRACT Accurate cropland maps are important input for various proposes, such as ecosystem service and land cover change monitoring, and the representativeness of sample training samples influence the cropland mapping accuracy significantly. This study aims to propose a new sampling workflow based on unsupervised cluster and multi-temporal interpretation (UCMT) for cropland mapping. The monthly composited image time series were unsupervised clustered using the Iterative Self organizing Data Analysis (ISODATA) and optimal temporal phases were selected using the Gini importance score calculated from Ramdom Forest (RF). Training samples for each cluster were generated and visually interpreted for the optimal temporal phases. The cropland of each temporal phase was identified using the corresponding training samples, and the cropland maps were generated by merging multi-temporal cropland results. Results in two study regions showed that training samples generated using UCMT had good potential to identify cropland with overall accuracies higher than 94% in both study regions. In addition, comparing with randomly generated training samples, UCMT samples were less affected by training sample size as Producer’s accuracies and User’s accuracies were higher than 80% when 100 training samples used.