Automated mapping of rice fields using multi-year training sample normalization

Rice agriculture is of great ecological, environmental, and socioeconomic importance in the Lower Mississippi Alluvial Valley, as its distribution and size heavily impact food production and a number of ecosystem services. Long-term rice mapping is challenging as a result of insufficient training data - both in spatial amount and in temporal coverage, the high cost of powerful geospatial data processing platforms, and incomplete image coverage during the critical window to capture the unique rice signals. Here, we developed a simple yet effective method for rice field extraction without heavy reliance on the complete profiles of Landsat time series or repeated training data. The core is a multiple-year training sample normalization that extends the samples obtained in one year for classification in another year. Pseudo-invariant objects and a set of linear regressions were used to predict what the given vegetation index values of training samples would be if they had been acquired under the same conditions in a different mapping year. The generated pseudo training samples were further utilized to classify the mapping image. We experimented with four years' Landsat Thematic Mapper and Operational Land Imager data and achieved comparable accuracies as the single-year classification. Because of its simplicity and low computational requirements, it can be efficiently implemented on cloud computing platforms, such as Google Earth Engine platform. This technique provides an affordable and effective solution to derive crop distribution information on a large-scale basis.