Optimizing the number of training areas for modeling above-ground biomass with ALS and multispectral remote sensing in subtropical Nepal

Remote sensing-based inventories of above-ground forest biomass (AGB) require a set of training plots representative of the area to be studied, the collection of which is the most expensive part of the analysis. These are time-consuming and costly because the large variety in forest conditions requires more plots to adequately capture this variability. A field campaign in general is challenging and is hampered by the complex topographic conditions, limited accessibility, steep mountainous terrains which increase labor efforts and costs. In addition it is also depend on the ratio between size of study area and number of training plots. In this study, we evaluate the number of training areas (sample size) required to estimate AGB for an area in the southern part of Nepal using airborne laser scanning (ALS), RapidEye and Landsat data. Three experiments were conducted: (i) AGB model performance, based on all the field training plots; (ii) reduction of the sample size, based on the ALS metrics and the AGB distribution; and (iii) prediction of the optimal number of training plots, based on the correlation between the remote sensing and field data. The AGB model was fitted using the sparse Bayesian method. AGB model performance was validated using an independent validation dataset. The effect of the strategies for reducing the sample size was readily apparent for the ALS-based AGB prediction, but the RapidEye and Landsat sensor data failed to capture any such effect. The results indicate that adequate coverage of the variability in tree height and density was an important condition for selecting the training plots. In addition, the ALS-based AGB prediction required the smallest number of training plots and was also quite stable with a small number of field plots. (C) 2016 Elsevier B.V. All rights reserved.