Enhancing forest inventory Accuracy: Comparing 3D-CNN and k-NN with genetic algorithm Approaches using ALS data across boreal bioregions

• Total growing stock estimates of 3D-CNN are always more accurate than k-NN. • Tree-species specific volume estimates of 3D-CNN are mostly more accurate than k-NN. • In terms of bias no clear winner between tested methods. In 2020, the National Land Survey of Finland (NLS) started to acquire airborne laser scanning (ALS) data with a nominal pulse density of 5 pulse/m 2 , representing a significant increase compared to the ALS data collected in earlier scanning campaigns. In our previous study utilizing lower density ALS data we have found that a three-dimensional convolutional neural network (3D-CNN) yielded results comparable to our benchmark genetic algorithm supported k-nearest neighbors (k-NN) method. In this paper, we compared the performance of our 3D-CNN model using voxelized ALS data with an increased point density to the benchmark k-NN method in estimating forest stand variables. The emphasis of our study was on the generalizability of tested models trained on independent, spatially uncorrelated datasets. For this reason, we utilized three datasets from different parts of Finland, of which two (Mikkeli and Äänekoski) are situated in the southern and one (Kolari) in the northern boreal bioregion. Models were trained with each of the datasets and cross-validated using the two other datasets and results reported separately. The 3D-CNN outperformed our benchmark model by more than 9 % in terms of relative RMSE over all training sets, validation sets and forest variables. On the other hand, k-NN performed slightly better than CNN regarding average relative bias by 4.9 % over all datasets and variables. It Is also important to mention, that over 94 % of RMSE results were proven to be statistically significant, whereas over 40 % of bias results showed no statistically significant differences between the methods. Most remarkable differences are produced by models trained and validated by datasets from the same bioregion (Mikkeli and Äänekoski). Relative RMSE scores calculated from CNN predictions were lower compared to the benchmark method by 5.7 %, 25.3 %, 19.6 %, and 63.3 % for volume of total growing stock, pine, spruce, and broad-leaved, respectively. Based on these results we can state that models trained with the 3D-CNN used in our study are better generalizable at least within the same bioregion compared to our benchmark method. Increased prediction accuracy especially in terms of tree species-specific volumes could bring benefits to forest management practices, planning of harvest, or biodiversity research.