Mapping dryland biocrust and plant communities utilizing temporal signals from satellite imagery with deep learning

Mapping dryland photoautotrophs, including vascular plants and biocrusts, is critical to monitor and predict dryland alterations under a changing climate. However, methodologies to map these communities, particularly patches associated with developmental stages of biocrusts, are underdeveloped, largely owing to limited spatial and spectral resolution of satellite data. Near-daily multispectral imagery from PlanetScope (3-m) provides a promising data source to distinguish dryland biota. Complex temporal dependencies of biotic activity and stochastic rainfall in drylands require methods which can extract features common to biota exposed to intermittent rainfall. We present a novel deep learning approach based on long short-term memory (LSTM) neural networks to capture temporal dependencies of biotic activity from PlanetScope imagery series for mapping biocrusts (light cyanobacteria and dark-mixed types) and vascular plants in the Great Basin’s Desert Valley. Labels for model training and testing were generated from field surveys and very high resolution aerial imagery over 40 plots across the landscape in a wet (2022) and a dry year (2017). Our approach yields high accuracy in mapping biota at 3-m, with an overall accuracy of 0.78, and user’s accuracy ranging from 0.67 to 0.88 across classes; when aggregated to subplot (50-m) scale, predicted proportional cover for biota agreed with those derived from aerial imagery, with R 2 ranging from 0.90 to 0.97 and RMSE ranging from 0.07 to 0.14. Photoautotroph separation improved when utilizing more frequent reflectance data. Maps derived using this approach revealed broad dark-mixed biocrust loss and transition to a likely lower productivity cyanobacteria dominant state in the Desert Valley over a drought period. This deep learning framework using PlanetScope time series facilitates ecosystem monitoring, improving our understanding of dryland state change, and likely can be applied across broader landscapes even when limited by the spatial or temporal resolution of data.