|dc.description.abstract||With the spread of urban culture, urbanisation is progressing rapidly and globally. Accurate and update land cover (LC) information becomes increasingly critical for protecting ecosystems, climate change studies and sustainable human-environment development. It has been verified that combining spectral information from remotely sensed imagery and 3D spatial information from airborne laser scanning (ALS) point clouds has achieved better LC classification accuracy than that obtained by using either of them solely. However, data fusions can introduce multiple errors. To solve this problem, multispectral ALS developed recently is able to acquire point cloud data with multiple spectral channels simultaneously. Moreover, deep neural networks have been proved to be a better option for LC classification than those statistical classification approaches.
This study aims to develop a workflow for automated pixel-wise LC classification from multispectral ALS data using deep-learning methods. A total of six input datasets with a multi-tiered architecture and three deep-learning classification networks (i.e. 1D CNN, 2D CNN, and 3D CNN) have been established to seek the optimal scheme that lead to highest classification accuracy. The highest overall classification accuracy of 97.2% has been achieved using the proposed 3D CNN and the designed input dataset. In regard to the proposed CNNs, the overall accuracy (OA) of the 2D and 3D CNNs was, on average, 8.4% higher than that of the 1D CNN. Although the OA of the 2D CNN was at most 0.3% lower than that of the 3D CNN, the run time of the 3D CNN was five times longer than the 2D CNN. Thus, the 2D CNN was the best choice for the multispectral ALS LC classification when considering efficiency. For different input datasets, the OA of the designed input datasets was, on average, 3.8% higher than that of the classic input datasets. Results also showed that the multispectral ALS data is superior to both multispectral optical imagery and single-wavelength ALS data for LC classification. In conclusion, this thesis suggests that LC classification can be improved with the use of multispectral ALS data and deep-learning methods.||en