This study developed a canopy structure estimation model based on time-series vegetation spectral variables to detect vegetation recovery in large-scale landslides and compare the recovery rates between vegetation indices and canopy structure. The analysis showed that the machine learning model effectively simulates canopy structure, achieving an R² of over 0.9 between simulated and observed values, enabling predictions of vegetation structural changes across broad spatial and temporal scales. The recovery trajectories of vegetation spectral indices and canopy structure revealed high variability in successional progress, with only approximately 14% of the landslide surface expected to recover to a mature forest state. The recovery of vegetation indices highlights saturation effects, tending to overestimate recovery rates and suggesting that well-restored vegetation could reach a mature forest state within 15 years. In contrast, canopy structure could require several decades to centuries to fully develop. Thus, vegetation indices are suitable for assessing early successional stages, while long-term restoration monitoring must also consider structural changes. Integrating spectral and structural information will facilitate a more comprehensive evaluation of restoration dynamics.

When the positioning and orientation equipment on an unmanned aerial vehicle (UAV) is unavailable, visual positioning technology can be utilized to perform spatial resection using only conjugate points from images to derive the vehicle's exterior orientation. This study proposes a visual positioning workflow and addresses the issue of significantly reduced matching success rates when using deep learning models for feature point matching due to planar rotation between images. By incorporating data augmentation with random image rotations, feature points are extracted using a feature extraction model and then input into the matching model for learning. Additionally, the study introduces interpolation methods and learnable parameter methods to replace the feature descriptors used for matching with traditional feature descriptors, enhancing rotational invariance. After extracting and matching feature points from the images, conventional photogrammetry spatial resection can be used to solve for the six exterior orientation elements of the camera mounted on the vehicle, thus achieving vehicle positioning. With the proposed visual positioning workflow, the best achievable plane position error is 3 meters, and the best achievable attitude angle error is 1.3°.

Land use and land cover (LULC) maps are essential foundational data for various landscape planning and resource management applications. Convolutional neural networks (CNNs), a deep learning method, can automatically extract features from remote sensing imagery and efficiently generate LULC maps. In recent years, CNNs have emerged as a widely recognized technique for image classification. This study utilized Sentinel-2 satellite imagery to construct a CNN model with a seven-layer architecture for LULC classification and compared its performance with the random forest (RF) machine learning algorithm. The results indicate that the CNN model outperformed RF, achieving an overall accuracy of 89% and a kappa coefficient of 0.84, compared to 87% and 0.81, respectively. Among the nine LULC categories, most classifications reached acceptable levels, with the exception of grasslands, fallow rice fields, and agricultural facilities. Overall, these findings demonstrate the potential of combining CNNs with satellite imagery for large-scale LULC mapping.

The Hehuan Mountain area has extensive coverage of Yushan Cane, while neighboring areas at similar altitudes do not exhibit widespread distributions of Yushan Cane. Therefore, an unsupervised classification of surface cover was conducted on the Hehuan Mountain area using Formosat-5 satellite images, and the classification results were evaluated. This process was combined with a 20m DTM for terrain analysis. The study found that the altitude range where Yushan Cane is located aligns with previous research, primarily between 3100-3600m. It also revealed that areas with Yushan Cane have gentler slopes than those with Abies Zone. Furthermore, the analysis showed that Yushan Cane in the study area is mainly distributed on eastern and southeastern slopes, differing from previous studies that indicated a southern slope preference. This study found that Yushan Cane is distributed at elevations above 1800m in the Hehuan Mountain area, with a significant increase above 3000m, primarily on mountain peaks, ridges, and other gentler terrain. In summary, besides altitude, slope is a key factor influencing the distribution of Yushan Cane in the Hehuan Mountain area.