Beyond Risk: Predicting Tropical Deforestation Intensity Patterns with Regression-Based Fully Convolutional Neural Networks

Tropical forests provide vital ecological, economic, cultural and climate-regulating services to local and global communities. However, these ecosystems are threatened by deforestation, often driven by complex and region-specific factors. Numerous studies have been conducted to predict the spatial distribution of deforestation risk, yet little research has explored the possible advantages of predicting deforestation intensity patterns. To support more effective forest management and conservation planning, this study examines the use of deep learning for predicting the spatial patterns of deforestation intensity.

This research develops and evaluates a regression-based ResUNet architecture for predicting deforestation intensity patterns. 
The deforestation datasets are, in most cases, highly skewed and zero-dominated, which poses the first challenge since this can significantly affect the predictive performance of the regression model. Several loss functions have been evaluated to mitigate this effect. The results illustrate how the Tweedie loss performs best. Furthermore, with a Root Mean Squared Error (RMSE) of 0.00494 on all values and 0.0169 on non-zero values, the Tweedie ResUnet model consistently outperforms the baseline XGBoost regression model. 

To test the model's cross-regional generalizability, four tropical regions were selected, each located on a different continent and characterised by varying deforestation drivers and dynamics. The Tweedie-ResUNet architecture was trained and tested on each study area. The differences in performance could be explained by regional characteristics such as data quality, topography, and seasonal cloud cover. However, the results still demonstrate a strong potential for the model's applicability to other tropical regions. 

The overall findings of this study suggest that deep learning models can be utilised to offer valuable insight into spatial patterns of deforestation intensity.