Tropical forests are hotspots of biodiversity, play a central role in global climate regulation, and provide ecosystem services to over a billion people. Yet, they face continuous threats, with the tropics accounting for 90% of global deforestation. A range of policies, regulations, and incentives has been introduced to address this, but their effectiveness depends on accurate, detailed, and timely data. Three types of data are particularly important: where deforestation occurs, what are its driving forces, and where it is likely to happen next. While systems for detecting deforestation are well established, methods for attributing deforestation to its drivers and forecasting future occurrences remain limited.
This thesis proposes three deep learning-based approaches to address this gap, utilizing remote sensing data that provide high-resolution, spatially comprehensive, and frequently repeated observations of tropical forests over multiple years. A particular focus is on designing neural network architectures that exploit the inherent spatio-temporal structure of the data, and that can fuse data from multiple complementary modalities.
In Chapter 3, a spatio-temporal architecture is proposed to estimate forest loss drivers from satellite image time series. The results demonstrate the benefit of integrating multi-temporal data, and a temporal attention module reveals the different strategies the model learns to recognize different driver classes. The trained model is applied to classify forest loss drivers sampled across the entire tropics.
Chapter 4 focuses on recognizing individual agricultural commodities associated with deforestation. Satellite image time series are complemented by geographic coordinates and country-level production statistics, providing spatial context in cases where specific commodities cannot be distinguished from imagery alone. A flexible, Transformer-based architecture is proposed for the fusion of these heterogeneous modalities, and is shown to significantly improve the recognition of spatially clustered land uses.
Chapter 5 investigates the feasibility of high-resolution deforestation forecasting based on local spatio-temporal patterns. Focusing on the Brazilian Amazon, where forest clearing follows characteristic spatial dynamics, it proposes an autoregressive approach in which predictions from one year are used to forecast the next. Two model variants are developed for short-term and long-term forecasting, and performance is evaluated across different temporal horizons.
Overall, the thesis demonstrates the potential of deep learning for large-scale, automatic deforestation attribution and forecasting, and the benefit of integrating multi-temporal and multi-modal data. It also identifies several challenges, particularly the scarcity and spatial bias of training data for deforestation attribution, the conceptual limitations of current attribution frameworks, and the inherent uncertainties of long-term forecasting,
