Predicting future patterns of land cover from climate projections using machine learning

Vegetation plays a crucial role in the Earths system, and its characteristics are strongly influenced by climate. Previous studies have investigated the climate-vegetation relationship, often attempting to predict vegetation types based on climate data. Many of them have utilized biome types as proxies for different vegetation forms. Biomes, although widely used, are not always optimal for this task. They are broadly defined, a priori linked to climate, and subject to change over time. This study proposes a novel approach by using the local composition of land cover (LC) categories as descriptors of vegetation types and examines the feasibility of modeling such compositions based on climate data. The investigation focuses on the Sahel region of Africa, which is tessellated into 5 x 5 km square tiles, serving as the basic units of analysis. The independent variable comprises a set of bioclimatic variables assigned to each tile, while the dependent variable consists of shares of each LC category within the tile. The modeling framework involves a set of n regressions, one for each LC category. The K-nearest neighbors (KNN) algorithm is employed to ensure that interdependently predicted shares sum up to 100%. The models performance is validated using 2010 data, where both climate and LC information are available. The mean absolute value of residuals ranges from 1% to 11%, depending on the LC category. Subsequently, future predictions of LC patterns are made for 2040, 2070, and 2100 using climate projections under IPCC scenarios 370 and 585. A novel visualization technique called synthetic landscape is introduced to visually compare the temporal sequence of predicted LC maps from 2010 to 2100 with similar sequences of biome maps and Koppen-Geiger climate type maps. This comparison highlights overall similarities across all sequences but also reveals some significant differences.