Implications of dissimilar geometries between training and population units in a three-dimensional deep learning–based forest inventory using airborne laser scanning and aerial image data

Remote sensing–based forest inventories create statistical relationships between 2D or 3D remote sensing data and field plots to predict forest attributes for grid-based population units that cover the entire area. In addition to conventional airborne laser scanning (ALS) feature-based models, deep learning alternatives have been extensively explored. However, most of the studies to date have only investigated performance in experimental setups where the training and testing data units have the same geometry (typically a circle). In practical applications, predictions are required for grid-based population units, which are typically square grid cells. While the conventional approaches that use lidar metrics appear to be largely unaffected by the geometric mismatch between the training and population units, this may not hold for the deep learning alternatives. Therefore, this study investigated the sensitivity of three-dimensional convolutional neural networks (3D CNNs) to discrepancies in shape and size between training and population units. We compared the predictions errors associated with 3D CNN architectures [fully convolutional (FCN) and Inception-V3] and Gaussian Process Regression (GPR), serving as a reference approach from conventional area-based approach, under different training (n = 376) and population unit configurations (n = 555 000). We used ALS data alone and in combination with aerial images to predict growing stock volume. Based on our results, both 3D CNN architectures were more sensitive than GPR to the discrepancy in shape and size between training and population units. For example, predictions to square population units using the FCN model trained with circular plots resulted in considerable underestimations (%RMSE: 49.8, %MD: −40.9). The utilization of either hexagonal or circular areas of remote sensing data in both training and population units resulted in the smallest prediction biases. Furthermore, we found that additional information on aerial images was beneficial with all experimented prediction methods. These results indicate that future work applying deep learning to model forest attributes with ALS data must consider shape discrepancies between training and population units.