A Constrained‐Regression Machine Learning Model for Predicting Future Natural Land‐Cover Composition From Climate Projections

Abstract

Predicting future vegetation patterns from climate change projections is central to assessing ecological impacts, but most existing climate–vegetation models predict coarse biome classes rather than fine‐grained land cover. Here we develop a constrained regression approach that predicts local land cover composition (LLCC)—the fractional cover of major natural vegetation categories within a fixed‐area tile—directly from local climate. Using ESA WorldCover land cover resampled to 100 m resolution and CHELSA bioclimatic variables at 5 km resolution, we represent the landscape as 5 km × 5 km tiles, each summarized by four bioclimatic predictors and a four‐part composition (tree cover, shrubland, grassland, and bare land). We train a ‐nearest neighbors regression model in which predicted compositions are obtained by averaging the compositions of the most climatically similar training tiles, which enforces the sum‐to‐one constraint by construction. Model performance is assessed by predicting present‐epoch LLCC from present‐epoch climate and comparing against observed compositions; mean absolute errors range from 1% to 11% across categories. We then forecast LLCC for 2040, 2070, and 2100 epochs under CMIP6‐based CHELSA projections for SSP 370 and SSP 585 scenarios and visualize the results with synthetic land‐cover maps constructed from present‐day patterns. The projections should be interpreted as climatically feasible compositions, reflecting climate controls but not explicitly accounting for land management, fire, soils, or other non‐climatic drivers.