Comparison of neural networks and k-nearest neighbors methods in forest stand variable estimation using airborne laser data

In the remote sensing of forests, point cloud data from airborne laser scanning contains high-value information for predicting the volume of growing stock and the size of trees. At the same time, laser scanning data allows a very high number of potential features that can be extracted from the point cloud data for predicting the forest variables. In some methods, the features are first extracted by user-defined algorithms and the best features are selected based on supervised learning, whereas both tasks can be carried out automatically by deep learning methods typically based on deep neural networks. In this study we tested k-nearest neighbor method combined with genetic algorithm (k-NN), artificial neural network (ANN), 2-dimensional convolutional neural network (2D-CNN) and 3-dimensional CNN (3D-CNN) for estimating the following forest variables: volume of growing stock, stand mean height and mean diameter. The results indicate that there were no major differences in the accuracy of the tested methods, but the ANN and 3D-CNN generally resulted in the lowest RMSE values for the predicted forest variables and the highest R2 values between the predicted and observed forest variables. The lowest RMSE scores were 20.3% (3D-CNN), 6.4% (3D-CNN) and 11.2% (ANN) and the highest R2 results 0.90 (3D-CNN), 0.95 (3D-CNN) and 0.85 (ANN) for volume of growing stock, stand mean height and mean diameter, respectively. Covariances of all response variable combinations and all predictions methods were lower than corresponding covariances of the field observations. ANN predictions had the highest covariances for mean height vs. mean diameter and total growing stock vs. mean diameter combinations and 3D-CNN for mean height vs. total growing stock. CNNs have distinct theoretical advantage over the other methods in complex recognition or classification tasks, but the utilization of their full potential may possibly require higher point density clouds than applied here. Thus, the relatively low density of the point clouds data may have been a contributing factor to the somewhat inconclusive ranking of the methods in this study. The input data and computer codes are available at: https://github.com/balazsan/ALS_NNs