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 dynamics. This paper presents Themeda, a modeling framework to predict land cover one or more years into the future, using artificial neural networks and time series of remotely sensed data from the world's largest intact savanna, across northern Australia. Themeda incorporates diverse spatio-temporal features, including 33 years of satellite-derived land cover, rainfall, temperature, fire scars, soil properties, and elevation, and generates a probability distribution for the future land cover for each pixel, across possible land cover classes. The model employs a ConvLSTM and a novel Temporal U-Net architecture, extending the U-Net with Long Short-Term Memory layers for multi-scale temporal processing. 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 for Food and Agriculture Organization Level 3 land cover classes and a Kullback-Leibler divergence of 1.65 × 10-3 for aggregated land cover predictions in 4000 m × 4000 m areas, surpassing baseline persistence models. The model performs strongly in predicting unseen test years, demonstrating robust generalizability. These probabilistic outputs and multi-scale temporal processing represent significant achievements for remote sensing applications, enabling improved ecological forecasting and supporting land use planning across diverse regions.

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

JEL Classification: Q57