Use of time-series L-band UAVSAR data for the classification of agricultural fields in the San Joaquin Valley

Abstract

The study of agriculture and agricultural production is an important part of societal planning for allocation of resources and economic forecasting. The prediction of crop output requires collecting accurate crop data throughout the growing season, in locations with a wide range of climatic conditions. Repeat observations from synthetic aperture radar (SAR) have been shown to be a reliable way of gathering frequent crop measurements, even in perpetually cloudy regions.In this work, repeat coverage of an area in California's San Joaquin Valley, with images taken by NASA's L-band Uninhabited Aerial Vehicle Synthetic Aperture Radar (UAVSAR) system approximately once a month during 2010 and 2012, was used to examine a time-history of backscatter signatures and polarimetric H/A/alpha (Entropy, Anisotropy, and Polarization angle) decomposition values of alfalfa, corn, and winter wheat with the objective of improving classification of individual crops. Distinguishable signatures were observed for all three crops. The signature was dominated by the growth stage and physical structure of the crops during the mature part of the growing season, and by weather events, planting practices, and harvesting procedures during other parts of the year. These data support previous findings that multiple images throughout the year capture the full growth pattern, allowing for more accurate identification of agricultural crops than what can be determined by a single image. The overall accuracies for this time-series classification were 75% using HV-polarized backscatter, 79% using the alpha decomposition, and 83% using the entropy decomposition, which were the three polarizations or decompositions with greatest separability between crops. It is shown that the full-year time series of images builds a more comprehensive model for crop backscatter than previous works.