kNN - Bagging NFI, GEDI, Sentinel-2 and Sentinel-1 data to produce estimates of forest volumes

This study presents a k-nearest neighbor bagging approach for predicting forest attributes, specifically growing stock volume (GSV), by integrating optical Sentinel-2, radar Sentinel-1, and spaceborne lidar GEDI datasets following a two-step procedure. Over a large region ( ∼ 4600 km 2 ), some selected GEDI variables were imputed to national forest inventory (NFI) plots, then models were developed to predict GSV based on the combined imputed GEDI and Sentinel variables. For GEDI imputation, three strategies were followed, each one using different auxiliary datasets: (A) solely relying on Sentinel data, (B) using Sentinel data complemented by maximum height values from both GEDI and NFI data sources, (C) using Sentinel and height values from an independent canopy height map derived from GEDI and Sentinel through Deep Learning. For each strategy, the models were built using a k-nearest neighbor approach and evaluated on an independent test dataset. The results indicated that the ability of Strategy A to predict forest volume as the percentage of explained variance ( R 2 ) was considerably limited (0.08) and the relative root mean squared error (RMSE%) was high (66%). Upon adding heights, Strategy B outperformed the previous strategy, with a R 2 of 0.58 and a relatively low RMSE% (43%). Compared to Strategy A, the use of estimated heights from a canopy height map improved predictions, but did not outperform Strategy B; R 2 was 0.34 and RMSE% 53%. Imputations yielded better models when using original GEDI footprint heights as inputs, rather than relying on height estimates extracted from the canopy height map. Therefore, small area model-assisted estimation may be preferred over wall-to-wall maps to reduce area-based local estimation errors.