Inspired by the success of deep learning techniques in dense-label prediction and the increasing availability of high precision airborne light detection and ranging (LiDAR) data, we present a research process that compares a collection of well-proven semantic segmentation architectures based on the deep learning approach. Our investigation concludes with the proposition of some novel deep learning architectures for generating detailed land resource maps by employing a semantic segmentation approach. The contribution of our work is threefold. (1) First, we implement the multiclass version of the intersection-over-union (IoU) loss function that contributes to handling highly imbalanced datasets and preventing overfitting. (2) Thereafter, we propose a novel deep learning architecture integrating the deep atrous network architecture with the stochastic depth approach for speeding up the learning process, and impose a regularization effect. (3) Finally, we introduce an early fusion deep layer that combines image-based and LiDAR-derived features. In a benchmark study carried out using the Follo 2014 LiDAR data and the NIBIO AR5 land resources dataset, we compare our proposals to other deep learning architectures. A quantitative comparison shows that our best proposal provides more than 5% relative improvement in terms of mean intersection-over-union over the atrous network, providing a basis for a more frequent and improved use of LiDAR data for automatic land cover segmentation.
This is an open access article distributed under the Creative Commons Attribution License
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited