Quantifying dwarf shrub biomass in an arid environment: comparing empirical methods in a high dimensional setting

• Published in: Remote sensing of environment Vol. 158; pp. 140 - 155
• Main Authors: Zandler, H., Brenning, A., Samimi, C.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.03.2015
• Subjects: Arid Environment; Aridity; artificial intelligence; Biomass; carbon; dry environmental conditions; Empirical analysis; Empirical modeling; Forests; Landsat; Landsat OLI; least squares; Mathematical models; monitoring; Multispectral remote sensing; prediction; RapidEye; Regression; Remote sensing; Sensors; shrinkage; shrubs; Tajikistan; uncertainty; Vegetation; vegetation cover; vegetation index; Tajikistan; Biomass; Arid Environment; RapidEye; Landsat OLI; Multispectral remote sensing; Empirical modeling
• ISSN: 0034-4257; 1879-0704
• DOI: 10.1016/j.rse.2014.11.007

Remote sensing based biomass estimation in arid environments is essential for monitoring degradation and carbon dynamics. However, due to the low vegetation cover in these regions, satellite-based research is challenging. Numerous potentially useful remotely-sensed predictor variables have been proposed, and several statistical and machine-learning techniques are available for empirical spatial modeling, but their predictive performance is yet unknown in this context. We therefore modeled total biomass in the Eastern Pamirs of Tajikistan, a region with extremely low vegetation cover, with a large set of satellite based predictors derived from two commonly used sensors (Landsat OLI, RapidEye), and assessed their utility in this environment using several suitable modeling approaches (stepwise, lasso, partial least squares and ridge regression, random forest). The best performing model (lasso regression) resulted in a RMSE of 992kgha−1 in spatial cross-validation, indicating that biomass quantification in this arid setting is feasible but subject to large uncertainties. Furthermore, pronounced over-fitting in some commonly used models (e.g. stepwise regression, random forest) underlined the importance of adequate variable selection and shrinkage techniques in spatial modeling of high dimensional data. The applied sensors showed very similar performance and a combination of both only slightly improved results of better performing models. A permutation-based assessment of variable importance showed that some of the most frequently used vegetation indices are not suitable for dwarf shrub biomass prediction in this environment. We suggest that predictor variables based on several bands accounting for vegetation as well as background information are required in this arid setting. •We model biomass with satellite data in a region with extremely low vegetation cover.•We apply six empirical models to test their ability in handling a large variable set.•Biomass quantification in arid environments is possible but uncertainties are large.•Lasso regression shows the best modeling performance.•Variables accounting for vegetation and soil background are required in this setting.