Nowadays, Unmanned Aerial System (UAS) imagery products also suffer from blurring and degradation caused by sun glint effects. Various techniques, including detection methods and specialized algorithms, are used to minimize sun glint's impact in aerial or remote sensing imagery. However, it remains uncertain whether the processing techniques used for low spatial resolution images can effectively be applied to images with high spatial resolution. By establishing the spatial relationships between the ground, sun, and sensor, a threshold for determining the presence of sun glint was established based on previously captured images, specifically for this research model. The findings of the results are presented from statistical, image-based, and physical spatial perspectives to identify the time period with the least sun glint during the target flight. This finding helps in reducing the effort required for sun glint removal. The key outcome of this approach is that employing photogrammetric techniques to establish a sun glint prediction model allows users to understand the distribution of sun glint throughout the entire image acquisition process during the planning phase. By adjusting the timing, it becomes feasible to plan flight schedules during periods of the day that offer higher efficiency in capturing useful images.

In recent years, deep learning has been used to construct quantitative indicators relevant to urban areas. Given the diverse array of dense billboards in Taiwan, this study aims to utilize deep learning techniques, including semantic segmentation and object detection, in conjunction with street view imagery to quantify the spatial distribution of signboards. Moreover, this study examines the spatial distribution patterns within the research area. The results demonstrate that the MIoU value of Deeplab v3+ model achieves 83%, while the Precision and Recall of YOLOv7 model achieves 91.7% and 87.1%. The analysis of spatial distribution patterns results align well with the actual distribution of billboards. This study can serve as a foundation for further exploration and application of billboards, as well as for integration with other related fields.

This study aims to investigate the impact of different algorithms on the temporal transferability of species distribution model (SDM) and the feasibility of using deep learning techniques to build SDM. The study focuses on Brainea insignis as the target species and utilizes two sets of samples collected with a 15-year interval. Experiments were conducted using the maximum entropy (MAXENT), random forest (RF), support vector machine (SVM), and U-net—a deep learning approach. The results indicate that MAXENT and SVM exhibit the best temporal transferability, while U-net also shows promising results. This highlights the research potential of deep learning, and future studies should consider incorporating a wider range of deep learning methods and continue experimentation. However, concerning environmental variables, relying solely on topographic factors may constrain the model's transferability, necessitating the identification of ecological factors with more direct causal relationships with the species to enhance transferability.

This study utilized unmanned aerial vehicle to obtain multispectral imagery of Mikania micrantha (bitter vine, BV) in two study areas, Plot A and Plot B. Employing maximum likelihood classification, random forest, and U-net, the study aimed to assess the model's performance in spatial extrapolation of spatial patterns of BV, with the goal of discovering new populations of previously unidentified species. Two sampling designs were employed for model training and testing, Set 1, conducting individual plot classification using training data from it’s owned; Set 2, performing spatial extrapolation from one plot with training data to another without training data, whose validation data there represent previously unidentified new populations. The results indicate that the three models performed well in Set 1. The kappa values for all models exceeded 0.75. However, the performance in plot B was slightly lower compared to plot A, possibly due to the more complex vegetation patterns in plot B. In Set 2, extrapolation from plot A to B showed fewer effective results compared to extrapolation from plot B to A. This discrepancy can be attributed to the differences in vegetation patterns and the flowering conditions of BV between the two areas. In terms of model performance, U-net demonstrated a better ability to capture the spatial patterns of BV, achieving the highest kappa value of up to 0.73 among the three models in Set 2. To comprehensively examine the model's performance in spatial extrapolation, future work will involve combining training data from both plots for extrapolation, and testing with other deep learning models such as convolution neural network (CNN) and environmental variables.