Forest canopy height estimation from ICESat-2 using TabPFN few-shot learning: a case study of the Auckland Region, New Zealand

Abstract

Forest canopy height (FCH) is a key parameter for characterizing forest vertical structure, stand development, and above-ground biomass. The spaceborne photon-counting lidar ICESat-2 provides along-track canopy height observations that can serve as training labels for regional FCH mapping, but its along-track sampling yields a limited number of usable training samples, and in complex forest landscapes the observations are affected by background noise, terrain relief, non-forest objects, and mixed pixels. Under such small-sample, highly heterogeneous conditions, conventional machine-learning models that rely on large-scale parameter tuning are difficult to fully exploit. This study introduces TabPFN, a prior-data fitted network for tabular data, into ICESat-2-driven FCH estimation over the Auckland Region, New Zealand, and builds a 30 m FCH estimation framework based on ICESat-2, multi-source remote sensing predictors, and TabPFN, with Random Forest, XGBoost, LightGBM, and CatBoost as conventional baselines. Segment-level canopy height labels (h_ph_98) were derived from ATL03/ATL08, and 9,824 quality-controlled training samples with 32 multi-source predictors were used. Using the Pearson correlation coefficient r and RMSE as primary metrics, TabPFN achieved the best accuracy in both the ICESat-2 internal test (r = 0.824, RMSE = 4.036 m) and the independent airborne-lidar external validation (r = 0.708, RMSE = 6.129 m), outperforming all four conventional models without iterative hyper-parameter tuning. All models exhibited severe underestimation in the tall-canopy segment, reflecting a common data limitation of passive-optical features. The results show that TabPFN few-shot learning provides a robust and tuning-free approach for FCH estimation from sparse ICESat-2 samples.