Advancing Yield Predictions in Pinus taeda (L.): An Artificial Intelligence (AI) Approach Leveraging LiDAR-derived Individual Tree Crown (ITC) Metrics, Competition Indices (CI), and Satellite Remote Sensing Indices

Abstract

This study assessed high-resolution UAV-LiDAR and multi-sensor satellite time series (Sentinel-1 and Sentinel-2) utilizing non-parametric machine learning and deep learning architectures across diverse planting densities (618, 1,236, and 1,853 trees ha-1) and thinning regimes to: 1) evaluate the accuracy of individual tree yield predictions over a 4-year interval using LiDAR-derived structural metrics and competition indices (see Chapter 1); 2) assess the temporal transferability of a single-date LiDAR acquisition for forecasting annual yield over a 7-year horizon (see Chapter 2); and 3) evaluate the efficacy of fusing continuous optical and synthetic aperture radar (SAR) time-series data using deep learning sequence models for plot-level yield estimation (see Chapter 3). Chapter 1 results showed that machine learning models significantly outperformed traditional parametric methods for medium-term yield prediction. Support Vector Machine (SVM) achieved the highest individual tree-level accuracy (normalized root mean square error (nRMSE) of 9.59%, R2 of 0.59) and underestimated stand-level volume by only -1.50%, while Random Forest (RF) achieved an nRMSE of 10.86% (R2 of 0.48) and overestimated stand volume by 1.53%. Chapter 2 results demonstrated the temporal transferability of a single age-8 LiDAR acquisition to predict annual growth from age 9 to 15. The RF model maintained high stability across the 7-year horizon (R2 greater than or equal to 0.83). Permutation feature importance revealed a biological shift where early-year predictions relied on individual structural metrics (tree top height increased MSE by 161.84%), while later-year predictions were dominated by distance-dependent competition indices (up to a 200% increase in MSE). Chapter 3 results showed that deep learning sequence models (Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM)) and RF successfully scaled predictions to the plot level (R2 of 0.49). GRU yielded the lowest overall error (RMSE of 60.38 m3ha-1, MAE of 34.28 m3ha-1). However, a distinct U-shaped error trend emerged across stand densities; the RF model achieved its lowest error in medium-density stands (618 to 1,236 trees ha-1; RMSE of 36.66 m3ha-1), while error rates sharply increased in high-density stands (greater than 1,237 trees ha-1; RMSE of 88.44 m3ha-1) due to signal saturation above 125 m3ha-1. Three primary conclusions come from this research: 1) non-parametric machine learning models utilizing individual tree crown metrics and competition indices accurately predict tree-level yield without violating the homoscedasticity assumptions that limit traditional linear models (see Chapter 1); 2) a single-date LiDAR acquisition is temporally transferable and captures fundamental ecological shifts from individual size-dependent growth to distance-dependent competition as canopy closure intensifies (see Chapter 2); and 3) deep learning architectures effectively fuse continuous SAR and optical satellite data for landscape-scale estimation, though sensor signal saturation remains a critical bottleneck in dense, mature plantations (see Chapter 3). From these findings, we hypothesize that deploying machine learning and deep learning models to integrate multi-scalar remote sensing data: 1) overcomes the spatial and logistical constraints of traditional plot-based field inventories, which 2) translates to highly accurate, continuous yield forecasting across varying silvicultural regimes, and 3) enables dynamic, data-driven precision forestry management, provided that sensor saturation thresholds in high-biomass stands are properly identified and mitigated.