Learning to Rank in 2D: Differentiable Spatial Prioritization for Deforestation Forecast

Environmental monitoring agencies often operate under strict intervention budgets, which makes the practical value of a forecast depend less on overall cell-wise accuracy and more on whether the model correctly prioritizes the small subset of locations that matter most. In this work, we reformulate deforestation forecasting as a spatial ranking problem and present a differentiable adaptation of probabilistic learning-to-rank to dense 2D prediction. The proposed framework combines a Top-K-oriented objective with a tournament-based sampling strategy that includes ground-truth injection to expose severely underestimated targets during training. Experiments on DETERB data from the Brazilian Amazon, using historical observations from 2018 to 2024 for training and a fully unseen 2025 period for testing, show that the proposed approach improves target prioritization and rank quality compared to conventional regression-based losses. Beyond deforestation forecasting, the formulation is applicable to dense prediction problems in which only a limited number of spatial targets can be inspected, reviewed, or acted upon.