Reconstructing Annual Dominant Tree-Species Distributions (1986–2024) in the Qilian Mountains via Multi-sensor Sample Transfer and Random Forest

Long-term tree-species information is fundamental for quantifying ecosystem services and forest climate resilience, yet multi-decadal mapping is often constrained by sparse field samples and heterogeneous satellite archives. Here we present a sensor-agnostic sample-transfer pipeline that reconstructs annual dominant tree-species distributions in the Qilian Mountains (north-eastern Tibetan Plateau) from a single-year field training dataset.

We harmonize optical missions (Landsat-5/7/8 and Sentinel-2) with multi-frequency SAR archives (ERS and Envisat C-band, Sentinel-1 C-band, and ALOS PALSAR L-band) and build a unified annual feature space combining spectral variables, SAR backscatter metrics, terrain predictors, and phenological descriptors. Phenology is derived from NDVI time series using Harmonic Analysis of Time Series (HANTS), yielding noise-robust seasonal metrics that remain comparable across sensors and years. To overcome the absence of historical labels, we transfer class labels from 6,268 field samples to each target year through a dual-constraint similarity screening: (i) feature-vector magnitude (Euclidean distance) and (ii) feature-vector direction (cosine similarity / spectral-angle-based measure). A thresholding rule discards ambiguous points and retains only reliable migrated samples. Annual maps are then generated using a Random Forest classifier (bagging and majority vote), while class imbalance is mitigated via downsampling and SMOTE.

Across nine sensor-integration periods spanning 1986–2024, the sample-transfer component remains stable despite changing sensors (transfer accuracy: 86.0–94.4%), and the resulting tree-species classification maintains consistently high accuracy (95.7–98.8%). Year-by-year assessments indicate overall accuracy and Kappa typically above 0.95 for most years and classes; performance reductions are mainly confined to rare taxa with limited observations. The final products provide consistent annual maps for six dominant tree genera (Betula, Juniperus, Picea, Populus, Pinus, and Larix) together with shrub–grass vegetation, cropland, water, and bare land, enabling robust quantification of multi-decadal changes in area fractions, spatial patterns, and centroid migration.

By coupling multi-sensor feature harmonization, HANTS-based phenology, and a dual-constraint sample-transfer strategy, this workflow offers a practical and generalizable route to recover multi-decadal tree-species dynamics from limited field data in mountain ecological barrier regions.