This study integrated 3S technology with artificial intelligence (AI), specifically utilizing Geospatial AI (Geo-AI), to spatially extrapolate the distribution of two invasive alien species (IAS), smooth cordgrass (SC) living in the intertidal zone and bitter vine (BV) emerging and growing inland in Formosa Island. The aim is to test whether three sampling schemes could improve extrapolation effectiveness and fill the gap in utilizing unmanned ariel vehicle (UAV) combined with deep learning in this field. Support vector machine, random forest, U-net, and DeepLabV3 were employed to recognize spatial patterns of SC and BV. Results indicate that the third scheme combining training data from two sampling sites in their own’s region yields optimal outcome, with deep learning outperforming machine learning, DeepLabV3 achieving F1 scores of 0.95 and 0.93 for SC and BV, respectively. This underscores the need for surveys with larger number of smaller separate sampling areas. Future works will utilize Geo-AI to search for additional IAS beyond the two regions, further realizing spatial extrapolation.

Spartina alterniflora (smooth cordgrass, SC) has invaded Formosa Island in recent years and is considered an invasive alien species (IAS). Careful assessment of IAS's impact on the island’s ecology is a top priority. Drones can improve the lack of autonomy and maneuverability of spaceborne and airborne systems, and better grasp plant’s phenology. We used drones with multispectral camera to acquire two phenological-phase orthoimages. Support vector machines (SVM), random forests (RF), and DeepLabV3 were used to extract SC from background, aiming to find out and map the spatial pattern of SC accurately. The results show that the two-date images can improve the accuracy; DeepLabV3 is far better than SVM and RF. The revalidation results show that the DeepLabV3 estimate was the closest to the real situation. The process done with the combination of drones and DeepLabV3 herein is effective and efficient, which can extract SC more accurately from background.

The purpose of this study is twofold, one is to establish species distribution model (SDM) of Brainea insignis (cycad fern, CF) based on the data measured by different GNSS receivers (survey-grade and recreational utility), so as to explore the impact of positioning quality on the model. Another one is to overcome the problem with insufficient samples for deep learning algorithm used to establish SDM. The results showed that even a few meters of error can cause the model to degrade, with the kappa value dropping by a maximum of 20.9%. Therefore, researchers cannot ignore the precision of species location intelligence. The way of extracting training samples has proved to be feasible to deep learning algorithm, in which one first establishes a high-potential habitat map by machine learning methods and then clips training areas input to this algorithm. Through this process, the model accuracy of U-net can reach above 90%. This opens up new possibilities for species distribution modeling.

Development of such invasive alien species (IAS) researches is challenged by grasping location intelligence (LI) of IAS in the large-scale environmental. In this study, multiple approaches were adopted to collect in-situ data, and four machine learning (ML), logistic multiple regression (LMR), support vector machine (SVM), maximum entropy (MAXENT), and random forest (RF), were used to develop SDMs for predicting IAS's patterns. In order to evaluate the key factors that affect the distribution of IAS, the study uses environmental variables including terrain-related and human-disturbance variables. The results show that the workflow for establishing SDM can overcome limitations in the natural environment with diverse survey approaches to find environment preferences for IAS. Moreover, we plot the prediction distribution with the best performing model, such as RF in this study and use this map to effectively limit the investigation range of IAS and improve the investigation efficiency.

Mikania micrantha (bitter vine, BV), an invasive plant, has remarkable phenological traits. Previous studies mostly used satellite or aerial photography systems that analysts do not have the right to control, also vulnerable to weather, causes the phenology of BV cannot be accurately photographed. Hence, drone was used to shoot four phases of images in this research. We simulated the classification results of the green leaf stage and flowering stage of BV could not be acquired. Support vector machine (SVM), maximum likelihood classifier (MLC), decision tree, neural network, and parallelepiped classifier were used, aiming to quickly identify BV's spatial pattern. The results show that SVM performs best, while MLC is the most efficient. Whether or not the green leaves and full flowers images of BV can be acquired will substantially affect the outcome, showing that the key point in monitoring BV is to grasp the two seasons, and only drone has the ability to do so.

As climate change getting severer, finding an effective way to establish specie distribution model (SDM) becomes vital, and collecting terrain-related variable is efficient. Chamaecyparis formosensis (Taiwan red cypress, TRC) grow around ridges above 1,800m covered with dry soil or gravel. Abies kawakamii (Taiwan fir, TF) grow near ridges and peaks over 3,000m with impermeable “terrain-shelterbelt” that block away overwhelming moisture. We used digital elevation models (DEM) with four different grid sizes (1, 5, 20, and 40m) to derive variables. We used several types of machine learning algorithms to model SDMs. The most accurate algorithms are decision tree and random forest, and the most suitable resolution is 5m. Elevation and topographic sheltering index are necessary for developing SDMs, and models can be more precise and accurate with using multiple layers of TSIs. As for different species, TRC tends to be majorly influenced by main high ridgeline, and TF is mostly affected by micro terrain.

Global Navigation Satellite System has a wide range of applications today, generating data with “location intelligence” that connects the physical and digital earth. In this study, different areas such as the campus of Chung-Hsing University, coastal zone, flat lands, and mountainous areas, two measurement methods, real time kinematic (RTK) and post-processing kinematic (PPK), were implemented by Trimble R12. Besides, Garmin 64st and six different brands of mobile phones were used for single-point positioning, then compared with the measurement results of R12 by calculating the root-mean-square of coordinates to understand the difference in positioning quality. In the results, as terrain becomes complex and steep with increasing altitude, the Position Dilution of Precision (PDOP) of R12 increased from 1.37 to 3.18, and the average root-mean-square deviation rose from 3.85 to 21.50 m. It can be seen that terrain together with the altitude has a great influence on the positioning accuracy and then will have different effects on the quality of the obtained location information.

Taiwan red cypress (TRC) forests generally grow in the fog-forest belt (FFB) with elevation above 1,800m. In contrast to Taiwan fir (TF) growing in depressions like a shelterbelt but impermeable, TRCs can occur at or near dry soils-and-gravels covered ridges or peaks where their leaves can intercept rich moisture as water. Hence, the study reversely cast from TF case to figure out why TRC forests are absent in the FFB areas of Huisun Experimental Forest Station (HEFS) in central Taiwan. It attempted to build species distribution models (SDMs) for TRCs and assess reversely the positive effect of topographic shelters associated with TFs on TRC forests. The decision tree (DT) and maximum entropy (MAXENT) models were developed with ecological parameters, including elevation, slope, aspect, global solar radiation (GSR), normalized difference vegetation index (NDVI), and topographic sheltering index (TSI). Modeling outcome indicates that topographic variables currently used in SDMs, except elevation and TSI, are useless for explaining the spatial patterns of TRCs. The statistics of TSI values are much lower at the TRC forests than those at the background areas. From the distribution TRC forests in the HEFS, this station lying in the occlusion of the northeast monsoon (NEM) in winter laden with rich moisture can be concluded. This outcome indicates why TRC forests cannot grow in the FFB areas in HEFS with elevation above 1,800 m. Consequently, TSI included in SDMs is useful to improve the predictive ability of SDM. However, the TSI will need to consider progressive lessening of NEM in winter with southward latitude as the NEM blows from northern Taiwan, through the central island, and to southern Taiwan. Moreover, a future study will confirm that the HEFS is in the occlusion of the southwest monsoon in summer from its southwest.

Species distribution model (SDM) has been the core of spatial ecology and it can provide a measure of a species' occupancy potential in areas not covered by biological surveys and consequently is becoming an essential tool to forest management. This study developed a framework for modeling two representative tree species in central Taiwan. The SDMs based on topographic variables and vegetation index derived from SPOT and WorldView-2 images for predicting potential habitat of two tree species in a GIS by using maximum entropy, DOMAIN and BIOCLIM). The results showed the variance in model accuracy across species was greater than that across techniques. Besides, SDM models merely based on topographic variables and sample distributions corresponding to them could not be applied on a larger spatial scale. More importantly, Adding spectral variable might offer high potential value, improving model prediction on a small and large spatial scale especially using Worldview-II image data.

Ecological niche modeling (ENM) coupled with 3S has become increasingly important for environment monitoring. Chamaecyparis formosensis (Taiwan red cypress, TRC) only grows in Huisun's Shou-Cheng Mountain. We used GIS to overlay physiographic variables and vegetation index with TRC samples. We developed ENMs by using generalized linear model (GLM), maximum likelihood (ML), maximum entropy (MAXENT) and BIOCLIM. Results indicated that the accuracies of four models increased linearly with the sample size and also the mean kappa value of SPA methods (0.95) was better than that of SPO methods (0.85) for predicting the suitable habitat of TRC forests in the study. The variables used except elevation could not reflect the relationship between humidity and topographic sheltering characteristics (wind). Hence, we proposed a hypothesis: northeastern seasonal wind with humidity cannot fully blow into Huisun due to its topographic sheltering effects. We will attempt to incorporate proxy indicators of wind and humidity into models.