Landcover Classification Using Radar Observations from the SMEX02 Campaign

Land classification using radar measurements has the potential to provide high spatial resolution information of Earth's land surface. In particular, the upcoming combined L- and S-band measurements of the NASA/ISRO Synthetic Aperture Radar (NISAR) mission have a clear potential for providing land surface classification information. Radar-derived land classification information can also support the simultaneous retrieval of other land surface geophysical information (e.g., soil moisture) from radar measurements. Existing land surface classification databases are typically available at spatial resolutions of 100 m or coarser or may only have very sparse temporal resolution. Radar-derived classifications used jointly with other radar-based geophysical retrievals in contrast provides information at the time and location observed by the radar system. Knowledge of specific crop types is beneficial not only for the retrieval algorithms but also a valuable product for agricultural monitoring and food security.

This study investigates the use of L- and S-band radar measurements in multiple polarizations for land surface classification using radar observations from the SMEX02 campaign in central Iowa. The classification approach applies the MATLAB Classification Learner (Weighted KNN) toolset, and classification performance is studied for both single-frequency and multi-frequency neural networks utilizing (HH and HV) or (HH, VV, HV, and VH) polarized backscatter. The results generally show higher accuracy in classifying corn versus soybean fields and show that approaches utilizing all polarized combinations outperform models utilizing only (HH and HV). Overall, the largest improvement in classification accuracy occurs when both L- and S-band data are included. The final performance results show an accuracy of 96.55% or greater for corn field detection and 84.83% or greater for soybean field detection. These results show the potential for land classification using radar observations in upcoming missions and the potential impact that multiple frequencies can have on performance accuracy.