LAUSR

Abstract A framework for estimating aboveground forest carbon stock (AFCS) is required for measurement, reporting, and verification (MRV) systems under the United Nations Collaborative Program on Reducing Emissions from Deforestation and Forest Degradation (REDD) in Developing Countries. Recently, methods for estimating the spatial distribution of AFCS using remotely sensed datasets and multiple prediction techniques have been found to be useful, particularly given the prohibitive costs of acquiring the necessary sample sizes for sufficiently precise pure field-based estimation of this variable. The objective of the study was to assess and compare the capabilities of airborne laser scanning (ALS) data, L-band radar data and UltraCam images in combination with four commonly used prediction techniques for estimating mean AFCS per unit area: multiple linear regression (MLR) and the nonparametric k-Nearest Neighbors (k-NN), support vector regression (SVR), and random forest (RF) algorithms. Our study area was a part of Hyrcanian deciduous forests in two managed and unmanaged stands at Shast Kalateh forest. We used a systematic sample consisting of 308 circular field plots of 0.1 ha located at the intersections of a 150 × 200 m grid with a random starting point and remote sensing-derived metrics as auxiliary data. We used 67% of the sample plots for training purposes and the remaining 33% for validation. Also, we used the model-assisted estimators to statistically rigorously estimate mean AFCS per unit area and its standard error (SE). Among the remotely sensed datasets, considered singly, the ALS data with R2⁎ = 0.34, rRMSE = 47.42% and relative efficiency (RE) = 1.51 produced the greatest accuracy and precision for AFCS estimation. RE can be interpreted as the factor by which the sample size for the pure field-based estimator would have to be increased to obtain the same precision as for the model-assisted estimator using auxiliary data. Among the remotely sensed datasets considered in combination, the ALS and PALSAR dataset with R2⁎ = 0.41, rRMSE = 44.80% and RE = 1.70 produced the greatest prediction accuracy and precision and increased the proportion of variability explained relative to ALS and PALSAR separately by 7% and 36%, respectively. Contrary to our expectation, the combination of PALSAR and UltraCam data decreased the precision of AFCS estimates. All the remotely sensed datasets, singly and in combination, with the most accurate prediction techniques for each combination produced estimates of AFCS that were within the 95% confidence interval for the field-based estimate.