Multi-output prediction of global vegetation distribution with incomplete data

Rita Beigaite, Jesse Read, Indre Zliobaite

Abstract:

As the climate is changing, large changes in vegetation distribution are already taking place and expected to happen in the future. Our goal is to explore the links between climate and vegetation, and build a predictive model mapping climatic conditions to vegetation cover, based on global remote sensing data. The main challenge is that many areas in the world are already significantly impacted by human activities, and natural mosaic of vegetation is altered, which makes natural vegetation data incomplete and increasingly inaccurate after fractions of agricultural and urban activity are removed.
Here we employ multi-output feed-forward neural networks for predicting natural vegetation cover from local climatic conditions.
We conduct experiments to evaluate how accurate predictions of the vegetation fraction can be and how they are affected by human-altered observations.
Results show that it is possible to make such predictions with high accuracy even if the training data are incomplete.

Keywords: multi-output prediction, incomplete data, global vegetation modeling

TL;DR: Predictive modeling of global vegetation distribution from climatic conditions based on remote sensing data which is incomplete in a non-uniform way.