Deep Learning Classification of Cheatgrass Invasion in the Western United States Using Biophysical and Remote Sensing Data

Multi-Species Inference of Exotic Annual and Native Perennial Grasses in Rangelands of the Western United States Using Harmonized Landsat and Sentinel-2 Data

The invasion of exotic annual grass (EAG), e.g., cheatgrass (Bromus tectorum) and medusahead (Taeniatherum caput-medusae), into rangeland ecosystems of the western United States is a broad-scale problem that affects wildlife habitats, increases wildfire frequency, and adds to land management costs. However, identifying individual species of EAG abundance from remote sensing, particularly at early stages of invasion or growth, can be problematic because of overlapping controls and similar phenological characteristics among native and other exotic vegetation. Subsequently, refining and developing tools capable of quantifying the abundance and phenology of annual and perennial grass species would be beneficial to help inform conservation and management efforts at local to regional scales. Here, we deploy an enhanced version of the U.S. Geological Survey Rangeland Exotic Plant Monitoring System to develop timely and accurate maps of annual (2016–2020) and intra-annual (May 2021 and July 2021) abundances of exotic annual and perennial grass species throughout the rangelands of the western United States. This monitoring system leverages field observations and remote-sensing data with artificial intelligence/machine learning to rapidly produce annual and early season estimates of species abundances at a 30-m spatial resolution. We introduce a fully automated and multi-task deep-learning framework to simultaneously predict and generate weekly, near-seamless composites of Harmonized Landsat Sentinel-2 spectral data. These data, along with auxiliary datasets and time series metrics, are incorporated into an ensemble of independent XGBoost models. This study demonstrates that inclusion of the Normalized Difference Vegetation Index and Normalized Difference Wetness Index time-series data generated from our deep-learning framework enables near real-time and accurate mapping of EAG (Median Absolute Error (MdAE): 3.22, 2.72, and 0.02; and correlation coefficient (r): 0.82, 0.81, and 0.73; respectively for EAG, cheatgrass, and medusahead) and native perennial grass abundance (MdAE: 2.51, r:0.72 for Sandberg bluegrass (Poa secunda)). Our approach and the resulting data provide insights into rangeland grass dynamics, which will be useful for applications, such as fire and drought monitoring, habitat suitability mapping, as well as land-cover and land-change modelling. Spatially explicit, timely, and accurate species-specific abundance datasets provide invaluable information to land managers.