Towards high-resolution forecasting of tropical deforestation with deep learning

This repository contains the data and selected trained models used in “Towards high-resolution forecasting of tropical deforestation with deep learning”. The data archive structure: inputs/ elevation.tifRaster layer of elevation covering the study area deforestation_maps/Annual raster maps of primary deforestation for the years 2010–2020. lulc/Annual land use/land cover raster maps for the years 2010–2020. models_trained/ 7f9gc9px/, b7pa58ky/, dndvccf8/, efd6ara9/, v5kqobut/Each folder contains: best.pt: Saved model checkpoint. config.json: Configuration file used for training the model. inference/ inference_area.gpkgFile defining the spatial area used for inference. It approximately corresponds to the Amazon biome municipalities_amazon_biome/Shapefile of municipalities within the Amazon biome. predictions/Raster files (2016.tif, 2017.tif, 2018.tif) with MYFORE prediction outputs. results/ engelman_results.pklNumerical results used for comparison with Engelman et al. loss_fc.pklNumerical results for comparing loss functions input_combinations.pklNumerical results for experiments with different input covariates. sequence_length.pklNumerical results for comparing different input sequence lengths. temporal_generalization.pklNumerical results for evaluating temporal generalization across years. annotation.csv - one row per data sample; contains unique identifier sampleid polygon geometry and the upper left position within the input rasters (row, col) total number of deforested (positive) pixels across all years counts of deforested pixels from 2010 to 2020 dataset split (train, val, or test) -