Does Spatial Context Always Matter? Weakly Supervised Forest Disturbance Detection Across Europe

Forest disturbances are a major driver of change in European forests, yet mapping their timing remains difficult because reference data are noisy and spatially sparse. Labels typically exist for a single pixel in a patch over multiple years, while the surrounding neighborhood is unlabeled. In this work we formulate annual forest disturbance detection as a weakly supervised problem on Landsat composites over Europe, and analyze how much spatial context is needed when only the central pixel is supervised. We compare temporal-only, spatial-only, jointly spatiotemporal, and factorized spatial-then-temporal architectures. With labels at only one pixel per patch, a dense spatial disturbance map cannot be evaluated, so year of change is inferred from the predicted temporal trajectory at the labeled location. Our experiments indicate that spatial context does not improve correct year detection overall and can even reduce it. Temporal-only models perform best (F1: 0.77, AUPRC: 0.81), indicating that the more discriminative signal generally lies in the spatiotemporal trajectory. An exception is slow, gradual disturbances, where spatial context appears beneficial, likely because a weak per-pixel signal can be stabilized by neighboring pixels undergoing the same process. These results suggest that, under single-pixel supervision, additional spatial context is not required for dating most forest disturbances, but may help for slow-onset events.