Detection of inundated vegetation using ALOS-1 L-Band over 2006-2010: Testing over conterminous USA

Global maps of inundated wetlands are needed to estimate their contribution to global methane emissions but is challenged by the difficulty of detecting sub-canopy inundation. Flooded forests with thick canopies, such as the those in the tropics, are among the largest methane emitters as well as particularly challenging to delineate. L-Band SAR has emerged as one of the best tools for detecting flooded vegetation, but disentangling the overlapping non-linear effects of volumetric and double-bounce backscatter against varying combinations of vegetation, soil moisture, inundation and incidence angle remains challenging.

We use ALOS-1 L-band SAR imagery to delineate inundated vegetation across the conterminous USA and tropical test regions as a step toward expanding globally. We combine HH-polarized retrievals from different beam modes to increase temporal coverage between 2006-2010. Individual scenes radiometrically terrain-corrected by the Alaska Satellite Facility (n=446k) at a 30-m resolution are ingested into the Google Earth Engine where they can be analyzed alongside contemporaneous Landsat indices of vegetation and water cover. We trained a wetland classification model with manually-interpreted wetland polygons from the National Wetland Inventory. Because of the uneven acquisition coverage of ALOS, we developed a random forest classification model leveraging the seasonal inundation signal with harmonic models over regions of high revisit (>50 images). Individual images are then classified as inundated/non-inundated based on an unsupervised thresholding to estimate inundation duration. The resulting map of the wetland inundated state is compared to static and dynamic wetland maps for benchmarking. The outputs from ALOS classification can improve representation of tropical wetland in global passive-microwave products, as well as help improving the spatial downscaling of these data.