Allometry-informed estimation of leaf area index in mountainous coniferous forests using Sentinel-2 spectral and textural features

Leaf area index (LAI) is a key descriptor of forest canopy structure, yet its reliable estimation in mountainous coniferous forests is hindered by topographic heterogeneity, complex stand structure and canopy-background mixing. We developed an allometry-informed Sentinel-2 framework for plot-scale LAI estimation in the Altai Mountains using field measurements from 47 plots. The framework incorporated DBH 2 × TH, where DBH is diameter at breast height and TH is tree height, as a structural proxy for canopy development. This term was combined with Sentinel-2 vegetation indices (VIs) and grey-level co-occurrence matrix texture features (TFs). Relevance ranking and redundancy filtering retained 25 VIs and 25 TFs from initial pools of 67 and 40 variables, respectively. GOSAVI combined with blue-band homogeneity was the highest-ranked inner-loop predictor combination. Under repeated nested cross-validation, the complete VI–TF modelling procedure achieved a pooled outer-loop R 2 of 0.73, an RMSE of 0.31 and an MAE of 0.25. By comparison, the structural-only baseline yielded an R 2 of 0.45, an RMSE of 0.44 and an MAE of 0.35. Comparisons with six benchmark regression algorithms showed that VI–TF combinations consistently contained useful LAI-related information, although predictive performance varied among algorithms and feature pairs. UAV observations from 12 overlapping plots provided a supplementary high-resolution assessment of cross-resolution consistency. Because these plots were a subset of the main field dataset, the UAV analysis was not considered an independent validation. Within the sampled forest and environmental conditions, integrating allometric structure with Sentinel-2 spectral and textural information improved plot-scale LAI estimation. Further evaluation using independent plots, years, regions and forest types is required to establish the broader transferability of the framework.