Abstract
Identifying drivers of deforestation is crucial for developing targeted conservation and land management strategies, and satellite data provide a long time series of data to understand deforestation dynamics. However, the timing of imagery after forest loss may affect classification accuracy, and optimal timing may be different for different drivers. Studies of broad-scale drivers across large and pan-tropical regions have shown that using time series can improve driver classification from satellite imagery, but requiring multi-year information means waiting longer after forest loss to classify what drives it. Our previously introduced model, Cam-ForestNet, was developed to use single-date imagery to classify fifteen direct detailed deforestation and degradation drivers for Cameroon. Here, we test whether the overall and per-class classification performance of Cam-ForestNet can be improved by either using imagery taken longer after a forest loss event or by incorporating a greater number of images, with performance evaluated using macro-average and per-class F1 scores to enable broad comparability across different contexts. Combining data up to four years after forest loss leads to improved model performance overall (macro-average F1 score) and for nearly all individual classes (per-class F1 scores). The classification of degradation drivers and slow-growing plantation benefitted most by incorporating time series data. However, when comparing approaches using only a single image from different years after a forest loss event, images from the first year following an event performed best, both overall (macro-average F1 score) and for most classes (per-class F1 scores), offering a promising strategy for relatively fast analysis of deforestation and degradation drivers following forest loss. We conclude that whilst multi-year imagery is beneficial, relying on a single image from the first year after forest loss still provides valuable and timely insights into the nature of drivers of forest loss.