Estimating stand metrics in mountainous temperate forests using UAS and satellite imagery: multiple linear regression or artificial neural networks?

Abstract

Economically and environmentally sustainable utilization of forest resources requires monitoring, but extensive field surveys require, and the deployment of high-quality LiDAR sensors is not always economically viable. However, optical imagery collected either by satellite or an unmanned aircraft system (UAS) is available even in remote areas. We predicted stand volume (V), dominant height (h dom), number of trees (N), basal area (BA) and quadratic mean diameter (dq), using images collected by Landsat 8 OLI (L8), Sentinel-1, Sentinel-2 (S2) and a UAS in natural pure Scots pine stands located in Northern Türkiye, which as a mountainous forested region presents a challenging test case. The stand metrics were modeled using multiple linear regression and artificial neural networks (ANNs) trained using Bayesian learning methods. Highest performance was found using L8 data for V (R2 = 0.84, logRMSE = 0.34) and S2 data for N (R2 = 0.70, logRMSE = 0.34), BA (R2 = 0.78, logRMSE = 0.31), h dom (R2 = 0.82, logRMSE = 0.19) and dq (R2 = 0.83, logRMSE = 0.07). The optimal orientation and window size varied depending on the stand characteristic. The ANN performed well for V (R2 = 0.83, logRMSE = 0.38), N (R2 = 0.81, RMSE = 0.34), BA (R2 = 0.84, logRMSE = 0.27), h dom (R2 = 0.71, logRMSE = 0.25) and dq (R2 = 0.89, logRMSE = 0.19). Stand metrics can be predicted in topographically challenging areas to a high accuracy using openly available satellite data.