Nathália Teles
I'm a veterinarian with a background in animal science and currently a PhD student in Environmental Sciences. I work as a researcher and field lead at the Image Processing and GIS laboratory at the Federal University of Goiás. I'm also involved in the Global Pasture Watch Initiative as a member of Lapig's team.
Sessions
The primary objective of this study was to develop and evaluate different remote sensing techniques for mapping Other Wooded Lands (OWL), while also assessing the accuracy and uncertainties associated with classifying OWL class compared to forest and grasslands. Additionally, we aimed to design a scalable process for large-scale OWL mapping. As defined by the Food and Agriculture Organization (FAO), OWLs are areas with 5-10% tree canopy cover for trees reaching a height of 5 meters at maturity, or with a combined cover of shrubs, bushes, and trees above 10 percent. Also, OWLs must span a minimum land area of 0.5 hectares and exclude predominantly agricultural or urban land uses. Three diverse landscapes were chosen based on expert input, encompassing natural regions globally and representing the three main land cover classes of interest: forest, OWL, and grassland. The selected areas were (1) Cheringoma, Sofala, Mozambique; (2) Cerrado biome, Goiás, Brazil; and (3) Albacete and Jaén, Spain. For each Area of Interest (AOI), we selected a Sentinel-2 MGRS tile that entirely covered the area. A stratified random sampling approach ensured robust sample collection across all land cover classes within each scene, resulting in over 1.7 million samples per scene. High-resolution imagery from Google Earth/Bing was utilized for visual interpretation. The mapping utilized data from 2022, encompassing a six-month window before and after the year of interest (totaling two years). A total of 174 metrics were calculated on data from various sources to characterize land cover for OWL modeling. Data processing was conducted using Google Earth Engine (GEE), and a Random Forest algorithm was employed for OWL land cover modeling. The resulting maps exhibited a global accuracy of 74.5% (Mozambique) and 76.5% (Brazil), Spain is currently under analysis. In Mozambique, the producer accuracy for OWL was 42.4%, with omissions associated with grasslands and forests at 34.5% and 21.5%, respectively. For the Cerrado region, both user and producer accuracies were notably higher, at 71.6% and 74.7% respectively. Mapping results were combined with ICESat-2 satellite lidar, where available, to investigate the vegetation height and structure of land cover classes. Top of canopy heights, median heights, and percent forest cover decreased between forest, OWL, and grassland classes. This methodology offers a scalable approach for mapping OWLs, contributing to improved deforestation monitoring and environmental protection efforts.