LiS2Forest: a multi-modal and multi-temporal dataset for benchmarking operational forest analysis

We introduce LiS2Forest, a national-scale benchmark combining LiDAR point clouds from the French LiDAR HD campaign, Sentinel-2 time series, and National Forest Inventory (NFI) field measurements across 29,429 plots in mainland France. We further exploit LiDAR HD overlap zones, where adjacent acquisition blocks intersect under varying acquisition setups, to extract a large collection of unlabeled paired point clouds and a labeled overlap subset. The benchmark supports four tasks: regression of six forest structural attributes, forest type and phenology classification, self-supervised regression under limited supervision, and cross-acquisition plot matching. We establish untuned baselines for all four tasks and identify key challenges including stem density estimation, mixed-composition classification, and cross-phenology matching. We release all datasets to advance multimodal and self-supervised forest representation learning.