Airborne Laser Scanning (ALS) can efficiently acquire large-scale point cloud data with high accuracy, which has become the major data source for Taiwan Digital Elevation Model (DEM). When generating ALS DEM, a significant amount of manual editing is needed to ensure the ground point classification, which are later used for DEM interpolation. In order to alleviate the manual burden, this research proposed an artificial intelligence (AI) ground classification workflow based on the geometric features from the ALS data. The geometric features are calculated and orthogonally projected to compose a “feature image”, which was further used as the training data for UNet. Then, by back-projecting the image classification results, the ground point within the ALS data can be classified. Three example datasets, including city, county, and forest scenes, were examined. The results showed that, in terms of areal percentage, 85.5%, 94.6%, and 74.3% of the AI-derived DEM are within ± 20 cm of the QC-inspected DEM for city, county, and forest scene, respectively. We further suggested that the confidence value output from the AI classifier can be used as an adaptive parameter to facilitate manually point cloud editing. Different threshold can be devised for different scene.

Multi-temporal optical satellite imagery is widely used in vegetation restoration monitoring research, and various vegetation indices are used to evaluate complex restoration processes. In this study, Landsat imagery was used to observe the vegetation recovery after the landslide of the typhoon Morakot in the Shenmu Village watershed area in 2009. Meanwhile, the differences in restoration of different landslide scales and slope locations, and the important factors affecting restoration were analyzed. The result shows that the recovery trend of Normalized Differential Vegetation Index is faster than Normalized Burn Ratio. The recovery rate of large-scale landslides is significantly lower than that of small and medium-sized landslides, while the vegetation restoration at the top, middle and lower part of the landslides have similar trends. Regression analysis shows that residual vegetation is the most important factor that affect the long-term restoration, and the vegetation before disturbance and the distance to the forest are also important factors; topographic characteristics provide indirect effects on vegetation colonization and restoration.

Generally, the volume of trees is often estimated using the tree volume equations, and the process of establishing the volume equations requires the measurement of the felled timber segments, which is a destructive investigation that is not only time-consuming and laborious, but also loses valuable trees. Due to the rapid development of Ground-based LiDAR technology, the use of LiDAR point cloud to establish a three-dimensional model of forest trees for volume estimation not only avoids felling trees, but also can accurately measure and quickly calculate tree volume. This study aims to explore the application of lidar point cloud data in urban tree volume measurement, and verify whether the TreeQSM algorithm is suitable for establishing an urban tree volume estimation model and verifying accuracy in Taiwan. The results showed that there were no significant differences in chest height diameter, tree height and trunk timber volume between in-situ measurement, point clouds manual measurement and TreeQSM estimation, which confirmed that TreeQSM could indeed be applied to urban tree volume estimation and establish a well-adapted timber volume equations, which can conform to the estimation and use of urban tree volume, and become an important tool for urban tree measurement.

In recent years, steel structure bridges have become increasingly popular in Taiwan, but Taiwan is prone to corrosion, so a lot of costs must be invested in the maintenance of bridge corrosion detection. In this study, UAV and Mask R-CNN are used to conduct corrosion identification experiments, and the results show that this method can improve the shortcomings of traditional bridge detection. At present, there is no corrosion dataset suitable for this study in Taiwan, so this study has established a corrosion data set by itself; It is imported into Mask R-CNN for training, and through the experiment of this study, the best super-parameter configuration is selected, that is, “optimizer SGD collocation learning rate 1 × 10-3 ”, and the evaluation index results of model training are: Recall can reach 97.1%, Precision can reach 90.4%, mAP can reach 91.0% and mIoU can reach 89.0%; After the analysis of the identification results, it is found that the background noise in the image will affect the identification of corrosion.