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

Authors: LIANG, LU - University Of Arkansas; RUNKLE, BENJAMIN - University Of Arkansas; SAPKOTA, BISHWA - University Of Arkansas; Reba, Michele

Submitted to: International Journal of Remote Sensing
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 1/23/2019
Publication Date: 4/17/2019
Citation: Liang, L., Runkle, B.R., Sapkota, B.B., Reba, M.L. 2019. Automated mapping of rice fields using multi-year training sample normalization. International Journal of Remote Sensing. 40:(18):7252-7271. https://doi.org/10.1080/01431161.2019.1601286.
DOI: https://doi.org/10.1080/01431161.2019.1601286

Interpretive Summary: Among the various crops, rice is extremely important to Arkansas: it is valued at $2 billion annually and is important to the economy of the state. Current techniques used to estimate the area and location of rice fields have limited accuracy and intensive labor requirements to identify field locations. A simple and effective method to identify rice fields was employed and tested for accuracy. The satellite images can capture the differences among different crop species by their distinct reflectance in the electromagnetic spectrum. An earth observation technique using Google Earth Engine (GEE), an in-development, cloud-based platform providing access to petabytes of satellite imagery data for planetary-scale analysis, was used. Combining this massive database with the parallel computing power of Google's infrastructure facilitates quick and easy analysis of satellite imagery on the ecoregion scale. The outcome of this research included annual rice maps in two major rice producing counties of Arkansas – Poinsett and Lonoke. Generally, the findings also contributed an affordable and effective solution to derive crop distribution information over larger areas and longer time periods. This method will be particularly useful in regions where agricultural levels are high but systematic surveys and censuses are lacking.

Technical Abstract: 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. Pseudoinvariant 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.