Remote Sensing and Machine Learning as Contemporary Indigenous Knowledge and Cutural Praxis: Mapping Sagittaria lancifolia in South Florida and the Carribbean

Abstract

Mapping dominant vegetative species across coastal wetlands provides useful insights into traditional ecological and cultural assets, as well as the hydrological and biogeochemical responses to weather variability and shifts in local and regional climate regimes. These areas are often inaccessible given the vast remoteness, complexity, and extent of such landscapes. Remote sensing is considered a vital tool for accurately mapping vegetation communities in coastal wetlands, supporting indigenous traditional ecological knowledge and cultural preservation. The primary objective of this research was to use a random forest (RF) classification method on combined multi-source remote sensing and ancillary data to map an important wetland species, Sagittaria lancifolia, or ‘bull-tongue arrowhead’ in the coastal wetlands and Everglades of South Florida and the Northern Karst of Puerto Rico. Data sources include remotely sensed datasets from harmonized Landsat - Sentinel-2(HLS-S30/L30), Sentinel-1 synthetic aperture radar, NASA’s Soil Moisture Active -Passive Level 4 Soil Moisture product, and ancillary datasets such as digital elevation models, soil water content, and soil porosity. Observational data were obtained from the Global Biodiversity Information Facility and field surveys in South Florida and Puerto Rico. The RF classification method was then applied to extracted pixel values of occurrence or non-occurrence of S. lancifolia and yielded an overall accuracy of ~87% accuracy (~65% Kappa statistic) in South Florida and ~95% accuracy (85% Kappa statistic) in Puerto Rico, when compared against test data not used in model training. The RF method produced slightly better accuracy with Landsat 8 data compared to the Sentinel-2 data. The RF-based variable importance analysis showed that elevation and soil water content were the most important factors distinguishing the occurrence of Sagittaria lancifolia from other nearby vegetation species in these regions. The trained model was used to create maps of Sagittaria lancifolia across three study sites in South Florida and two study sites in Puerto Rico. Most classification errors resulted from commission errors, likely due to the variety of ecosystems that S. lancifolia occur in, and the surrounding vegetation that it may be found embedded within. Additionally, sampling data was not an exhaustive list of Sagittaria lancifolia occurrences in the study regions. Nonetheless, this research demonstrated the potential of using satellite data and machine learning techniques in mapping a target vegetation species of high cultural and ecological value, S. lancifolia, within wetlands. Future work will include a larger spatial extent and be enhanced by ground truthing (in-situ) fieldwork to collect observation data in Puerto Rico. Studies can use maps derived from this method to monitor the distribution of S. lancifolia, or other species of interest, which serve a critical role in blue carbon accounting for future economic development and environmental monitoring, as well as serving a role in cultural heritage preservation by better understanding the reproductive habitat distribution of the endangered coqui Eleutherodactylus juanariveroii.