25-year, quarterly land change maps of China's Loess Plateau reveal long-term and substantial water-induced soil erosion mitigation

Abstract. Unsustainable human activities have driven global ecological degradation. In China, decades of restoration policies have been implemented to reverse this trend in severely degraded regions with catastrophic soil erosion, transforming them into landscapes of ecological recovery. However, the evolution of soil erosion in these regions remains poorly quantified due to the absence of high-resolution, long-term, and high-frequency monitoring data. Here, to address this gap and provide a reliable spatiotemporal benchmark dataset, we conducted the first 10–30 m quarterly wall-to-wall land change mapping for China's flagship ecological restoration site: the Loess Plateau, based on the developed cross-temporal consistency-constraint deep learning framework. The dataset was generated using over 10 TB of Sentinel and Landsat imagery and documents land-cover dynamics across 100 quarterly time steps from 2000 to 2024, showing an overall accuracy of 81.44 % based on 40 000 annotated samples and 79.8 % for third-party validation sources. The resulting maps record pronounced land-cover dynamics, including forest expansion (+13 131 km2), cropland expansion (+28 095 km2), and bare land reduction (−65 029 km2) over the past decades. Furthermore, the produced dataset was combined with environmental factors to measure the 25-year water-induced soil erosion, where comparison with government survey data shows strong consistency, with a mean absolute error of 4.50 %. The dataset further illustrates that long-term ecological interventions have substantially reduced erosion intensity in the region by 30 % over the past 25 years, from 13.34 to 9.35 t (hm2 a)−1. Based on this benchmark, the long-term, fine-grained soil erosion becomes possible to estimate. The data-driven analysis indicates that current erosion is most severe in the central and southwestern Loess Plateau, and scenario modeling based on multiple factors suggests that optimized vegetation distribution – including grassland expansion and cropland-to-forest conversion – could potentially reduce future erosion intensity to 6.42 t (hm2 a)−1. This dataset provides a comprehensive benchmark for erosion mitigation in the Loess Plateau and its underlying drivers, providing critical insights for sustainable land management, ecological restoration, and policy development both in China and across fragile ecosystems worldwide. The land-cover maps and soil erosion maps are available at https://doi.org/10.57760/sciencedb.33656 (Cheng et al., 2026).