Active layer thickness (ALT) is a key indicator of permafrost change, with important implications for soil hydrothermal conditions, carbon feedbacks, ecosystem processes, and cold-region infrastructure. However, hemispheric-scale ALT mapping remains challenging because field observations are sparse and unevenly distributed, while thaw depth is controlled by complex and nonlinear environmental interactions. Here, we compiled 2,196 annual ALT observations and developed an ensemble spatiotemporal machine-learning framework to reconstruct a continuous annual ALT dataset at 1 km resolution for the Northern Hemisphere permafrost region from 2000 to 2024. Spatial leave-one-site-out cross-validation yielded an ensemble <em>R<sup>2</sup></em> of 0.76 and an RMSE of 60.48 cm. Independent temporal evaluation showed a significant correlation between observed and predicted Sen’s slopes (<em>r</em> = 0.73, <em>p</em> < 0.001), with 86.8 % agreement in trend direction. The dataset reproduced broad latitudinal and elevational patterns, biome-related differences, and interannual variability. Comparisons with two existing 1 km hemispheric products showed that our reconstruction occupied an intermediate position in hemispheric mean ALT while preserving relatively fine spatial variability. Substantial inter-product differences in both magnitude and spatial structure further highlighted persistent structural uncertainty in large-scale ALT mapping. Pixel-wise uncertainty maps further provide spatially explicit information on prediction confidence. This dataset provides a spatially detailed, temporally continuous, and uncertainty-explicit resource for regional- to hemispheric-scale studies of permafrost dynamics, carbon-cycle feedbacks, ecosystem and hydrological responses, and infrastructure exposure.