Themeda: Predicting land cover change using deep learning

Abstract

Accurate land cover change prediction is vital for informed land management, and deep learning offers a flexible solution capable of capturing complex ecological patterns. This paper presents Themeda, a modeling framework using artificial neural networks that predict land cover category probability distributions based on historical data. It integrates ConvLSTM and a novel Temporal U-Net architecture, extending the U-Net with LSTM layers for multi-scale temporal processing, enabling fine-grained local and broader spatial pattern capture. Leveraging 33 years of historical data from the world's largest intact savanna, Themeda incorporates diverse spatio-temporal features like rainfall, temperature, elevation, soil types, land use, and fire scars. Themeda overcomes limitations of current spatio-temporal models by processing temporal data at multiple spatial scales, capturing local and regional ecological changes effectively. It achieves a 93.4% pixel-wise validation accuracy on FAO Level 3 classes and a KL divergence of 1.65e-03 for aggregated land cover predictions in 4000 m x 4000 m areas, surpassing baseline persistence models. The model maintains high performance on unseen test years, demonstrating robust generalizability. The probabilistic outputs and multi-scale temporal processing have significant implications for enhancing cellular automata and land use planning models and could be adapted for ecological forecasting in other regions.

Keywords: deep learning, neural networks, spatiotemporal prediction, land cover change, savanna

JEL Classification: Q57