Swedish forest-mapping AI works in Latvia with minimal retraining
Researchers successfully transferred machine learning models trained on Swedish forest data to Latvia, requiring little to no local calibration. The finding suggests AI-powered forest monitoring could scale across borders and regions, reducing costs for timber companies, carbon accounting, and forest management across Northern Europe.
Originaltitel: Transfer of forest attribute models from Sweden to Latvia using a DeepLab-XGBoost MetaPredictor
<p>Airborne laser scanning (ALS) and area-based modelling enable operational mapping of forest structural attributes, yet cross-border model transfer remains under explored. We assess whether a Swedish-trained "MetaPredictor"based in DeepLab v3 feature extraction from 9 & times; 9 & times; 60 ALS raster patches fused with tabular metrics via XGBoost, can be transferred to Latvia with minimal local calibration. Swedish training relied on ALS raster stacks aligned to National Forest Inventory (NFI) plots. Targets included mean height, mean diameter, basal area, stem volume, and above-ground biomass (AGB). Latvian evaluation used ALS point clouds processed with lidR to produce equivalent 60-band rasters and plot-level reference attributes. Latvian inputs were standardized using Swedish scalers, passed through DeepLab v3 (ResNet-101) to extract spatial features, concatenated with scaled ALS laser metrics, and evaluated under two regimes: zero-shot transfer (no local labels) and few-shot fine-tuning (1%-20% Latvian plots). Performance was quantified with RMSE, MAE, relative RMSE (rRMSE), and R2. In zero-shot mode, height transferred well (R2 = 0.91, RMSE 2.60 m, rRMSE 14%), diameter moderately (R2 = 0.60, RMSE 7.72 cm), while basal area and AGB showed weak to negative generalization (R2 = -0.07 and-2.29). Few-shot fine-tuning produced decisive gains: basal area RMSE dropped from 11.9 to 5.9 (R2 = 0.73), volume from 139.7 to 71.7 (R2 = 0.83), and AGB from 173.8 to 36.16 (R2 = 0.86), amounting to 48%-79% error reductions. Height and diameter, already transferable, improved incrementally. Results demonstrate that Swedish ALS-trained models can seed Latvian operational mapping with modest local calibration. The MetaPredictor design, combining deep spatial encodings with tabular metrics, supports rapid of forest attribute in with limited labelled data.</p>