Reconstructing canopy cover in a Cerrado–Pantanal ecological corridor from 1985 to 2024 with GEDI lidar, Landsat and neural network regression
Abstract Spaceborne lidar measures canopy cover only since 2019, which prevents reconstructing past vegetation structure. We tested whether canopy cover from the Global Ecosystem Dynamics Investigation (GEDI) can be extended backwards with Landsat and machine learning regression in the São Lourenço River Basin Ecological Corridor (765,910.91 ha), in the Cerrado–Pantanal transition, Brazil. Dry-season GEDI footprints of 2019–2021 were paired with 11 predictors from Landsat composites of the same year; multiple linear regression, random forest, gradient boosting and a multilayer perceptron were compared by spatial block cross-validation and tested on footprints of 2022. The selected model mapped five epochs from 1985 to 2024, change was assessed with a propagated-error threshold and a threshold sensitivity analysis, and the maps were compared with two independent tree cover products. The multilayer perceptron was selected (root mean square error, RMSE = 0.234), but all models performed within the variation between folds and compressed the cover range; the 2022 test gave RMSE = 0.248. Mean predicted cover ranged from 0.230 (2005) to 0.357 (2015), but the dry-season vegetation index at fixed training locations was higher in the Landsat 8–9 epochs than in the Landsat 5 epochs, so part of this variation reflects sensor and interannual effects. Above the propagated threshold, 6.5% of the corridor showed gain and 1.6% loss in 1985–2024, and almost no change exceeded the 95% threshold. GEDI-trained Landsat regression maps the spatial pattern of canopy cover consistently, but decadal change in this savanna requires cross-sensor calibration before it can be interpreted.
Authors
- David Albuquerque
Institutions
- Universidade Federal de Rondonópolis (BR)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-11653283/v1
- Primary Topic
- Remote Sensing and LiDAR Applications
- Type
- preprint