Distribution Statistics Preserving Post-Processing Method With Plot Level Uncertainty Analysis for Remotely Sensed Data-Based Forest Inventory Predictions

Abstract

Remotely sensed data-based models used in operational forest inventory usually give precise and accurate predictions on average, but they often suffer from systematic under- or over-estimation of extreme attribute values resulting in too narrow or skewed attribute distributions. We use a post-processing method based on the statistics of a proper, representative training set to correct the predictions and their probability intervals, attaining corrected predictions that reproduce the statistics of the whole population. Performance of the method is validated with three forest attributes from seven study sites in Finland with training set sizes from 50 to over 400 field plots. The results are compared to those of the uncorrected predictions given by linear models using airborne laser scanning data. The post-processing method improves the accuracy assessment linear fit between the predictions and the reference set by 35.4–51.8% and the distribution fit by 44.5–95.0%. The prediction root mean square error declines on the average by 6.3%. The systematic under- and over-estimation are reduced consistently with all training set sizes. The level of uncertainty is maintained well as the probability intervals cover the real uncertainty while keeping the average probability interval width similar to the one in uncorrected predictions.

Role of data uncertainty when identifying important areas for biodiversity and carbon in boreal forests

“…Variables were resampled simultaneously for each segment using a covariance matrix of errors. To avoid negative or unrealistically large values, and to maintain the covariance structure between variables, the segment samples were post-processed using the quantile matching procedure (Junttila and Kauranne 2018 ). For site type, data on the satellite-based site type and forest structural variables from Miettinen et al ( 2021 ) were used together with a probit model to calculate the probability of a segment belonging to each of the site type classes.…”