A river runs through it: Robust automated mapping of riparian woodlands and land surface phenology across dryland regions

• Published in: Remote sensing of environment Vol. 305; p. 114056
• Main Authors: McMahon, Conor A., Roberts, Dar A., Stella, John C., Trugman, Anna T., Singer, Michael B., Caylor, Kelly K.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.05.2024
• Subjects: algorithms; arid lands; automation; biodiversity; class; Classification; climate change; cloud cover; data collection; Drylands; energy; environment; Google earth engine; groundwater; highlands; humans; Landsat; landscapes; lidar; Phenology; radar; Random Forest; remote sensing; Riparian; rivers; spatial variation; stream flow; topography; uncertainty; watersheds; woodlands; Phenology; Riparian; Google earth engine; Drylands; Classification; Random Forest
• ISSN: 0034-4257; 1879-0704; 1879-0704
• DOI: 10.1016/j.rse.2024.114056

Riparian woodlands in drylands are critically important to human society, global biodiversity, and regional water and energy budgets. These sensitive ecosystems have experienced substantial degradation over the last several decades from climatic change and direct human activity. Nevertheless, quantifying long-term change in dryland riparian woodlands remains a major challenge, and much uncertainty exists in their remaining extent, historical breadth, and likely future trajectories. Dryland landscapes show large, fine-scale spatial heterogeneity in seasonal greenness patterns, driven in part by spatial variation in water availability. Riparian woodlands occur where water is concentrated in the landscape, either as aboveground streamflow or subsurface groundwater. In arid and semi-arid climates, this renders them phenologically distinctive from upland ecosystems. However, despite their importance and distinctiveness, there are currently no automated methods for delineating dryland riparian woodlands across regional extents in the cloud. Here we designed and implemented a cloud-based algorithm to retrieve dryland land surface phenology patterns from multispectral satellite imagery and conducted sensitivity analyses using real and simulated data to demonstrate that the approach is robust for MODIS, Sentinel-2, and Landsat over realistic ranges of noise and cloud cover. We then designed a series of random forest vegetation classifiers that integrate phenological and spectral information, vegetative structure from LiDAR, and topography from LiDAR or the Shuttle Radar Topography Mission. We implemented classifiers for three local study sites and then generalized our model to run regionally across the southwestern United States, with balanced accuracy for the riparian woodland class ranging from 94.5% to 97.5% when validated with local to regional datasets. Generally, phenological information proved more important than any other data source for mapping riparian woodlands, which showed more stability in interannual phenology than did upland vegetation types. To our knowledge, ours is the first regional, annual, automatically-generated and updated approach for mapping dryland riparian woodlands in the southwestern United States, paving the way for improved modeling and management efforts on watershed to regional scales. We also provide one of the first operational, exclusively cloud-based methods to extract dryland land surface phenology patterns using Landsat, Sentinel-2, MODIS, or other sensors, providing a framework for future studies investigating other aspects of long-term or spatial variation in dryland vegetative seasonality across the globe. •We developed rapid, cloud-based maps of phenology from multispectral sensors.•We validated our product using of cross-sensor comparisons and simulations.•We used random forest to map riparian woodlands using greenness phenology.•Maps had high accuracy for riparian woodlands across regions and years.•We extended our classifier regionally in Google Earth Engine.