Predicting global potential natural vegetation with an image recognition AI

Abstract:

Potential natural vegetation (PNV) is the vegetation cover equilibrium with environmental condition, which would exist at a given location without human land-conversion. For operational mapping of PNV, we developed an empirical model using a Convolutional Neural Network (CNN), which was trained by an observation based PNV map (Figure 1) and graphical images of global air temperature and precipitation at 0.5 degree resolution. The trained model well reconstructs an observation based global PNV map, demonstrating that this way of CNN application can capture empirical relationships between PNV and climate. Then, the trained model was applied to projected climate at the end of the 21st century, predicting significant shift of global PNV distribution with rapid warming trends (Figure 2).