Abstract

Satellite interferometric synthetic aperture radar (InSAR) is emerging as a viable low-cost alternative method to airborne laser scanning (ALS) for forest inventory though little research has examined its efficacy for plantation forests located in temperate regions on steep terrain. InSAR and ALS data were collected from Geraldine Forest which is located on rolling to very steep topography in the southeast of New Zealand. These data were combined with an extensive set of plot measurements from which mean top height (H), basal area (G), stem density (N), and total stem volume (TSV) were calculated. InSAR and ALS-based Random Forest models of each variable were developed and compared.Using the ALS data as a reference, the mean RMSE of the InSAR DSM and DTM surfaces were, respectively, 4.58 and 8.09 m and these errors increased to mean values of, respectively, 6.02 and 10.17 m for slopes of 40-50 degrees.ALS-based models were substantially more precise than those developed from InSAR for H (R-2 = 0.86 vs. 0.60; RMSE% = 5.47 vs. 10.8%), G (R-2 = 0.56 vs. 0.32; RMSE% = 21.5 vs. 30.4%), N (R-2 = 0.47 vs. 0.09; RMSE% = 32.3 vs. 43.2%), and TSV (R-2 = 0.70 vs. 0.41; RMSE% = 19.4 vs. 30.7%). The base metrics (i.e. ALS height and canopy cover variables) accounted for most of the variance in the ALS models with addition of further metrics providing <= 1% reduction in the RMSE%. Base metrics (i.e. InSAR observables) also accounted for most of the variation in InSAR models. Addition of metrics from a mixed Canopy Height Model (CHM), derived from InSAR Digital Surface Model (DSM) and ALS Digital Terrain Model (DTM), resulted in reductions in RMSE% of 3.1-5.4% for H, G, and TSV models with addition of textural metrics providing further reductions of 0.2-0.3% for H and G models. Addition of metrics from the radar CHM, derived from the InSAR DTM and DSM, and texture metrics reduced the RMSE% of the base model for N by 2.3% and 0.5%, respectively.The results were generated using a SAR image pair with a height of ambiguity (HOA) that was higher than ideal, which reduced the sensitivity of results to changes in terrain. Despite this limitation, and the steep slopes throughout the forest, the InSAR models described here had comparable precision to developed InSAR models for key inventory metrics from previous studies.

Keywords

Forest inventory; LiDAR; Machine learning; Random forest regression; SAR interferometry; Synthetic aperture radar; TanDEM-X; Volume

Published in

Remote Sensing of Environment
2020, Volume: 246, article number: 111833
Publisher: ELSEVIER SCIENCE INC

SLU Authors

• Persson, Henrik

  • Department of Forest Resource Management, Swedish University of Agricultural Sciences
    

Sustainable Development Goals

SDG15 Protect, restore and promote sustainable use of terrestrial ecosystems, sustainably manage forests, combat desertification, and halt and reverse land degradation and halt biodiversity loss


UKÄ Subject classification

Remote Sensing


Publication identifier

DOI: https://doi.org/10.1016/j.rse.2020.111833

Permanent link to this page (URI)

https://res.slu.se/id/publ/106714