Human Modification Index forecasts from 2025 to 2040

Global probabilistic forecasts of the Human Modification (HM) index at 1 km resolution and 5-year intervals from 2025 to 2040. The forecasts come from a convolutional LSTM trained on the harmonized 1990–2020 HM time series of Theobald et al. (2025). The model also uses dynamic covariates (HM component threats, gridded population, gridded GDP) and static covariates (CHELSA climate, EarthEnv elevation, WDPA protected-area masks and a development suitability index). It makes independent predictions for four horizons (2025, 2030, 2035 and 2040). For each horizon it estimates the full distribution of possible HM values at every pixel. This record contains single-band float32 GeoTIFFs in EPSG:4326 with ~1 km pixels, covering land between 84.00°N and 70.00°S. There is one raster per forecast year for each of five variables: - the mean HM- the lower bound (2.5th percentile)- the upper bound (97.5th percentile)- the expected probability that HM will exceed 0.1- the expected probability that HM will exceed 0.4 Files are named hm_YEAR_VARIABLE.tif, where YEAR is 2025, 2030, 2035 or 2040 and VARIABLE is lower, mean, upper, gt01 or gt04. HM values range from 0 (unmodified) to 1 (completely modified). The record also includes the historical HM observations for 1990–2020, resampled and masked to the forecast grid and named hm_observed_YEAR.tif, so you can work with one continuous 1990–2040 time series. These GeoTIFFs are summaries of a larger product. The full probabilistic forecasts are on the Earthmover Data Marketplace as the Arraylake repository the-nature-conservancy/hm_probability_forecast. They give the complete predicted distribution of HM for every pixel and forecast year, as 64 quantile levels from 0.025% to 99.99%. From these you can derive any percentile, the expected HM, or the probability of crossing any HM threshold, not just 0.1 and 0.4. The marketplace version also includes the observed 1990–2020 HM series on the same grid.