Research in Detection and Evaluation of Green Spaces Using UAV and Artificial Intelligence

Abstract Numerous efforts have been made to evaluate green areas, ranging from manual methods like tree and leaf counting to more modern approaches such as using satellite imagery. However, these methods have limitations, including high costs, resource-intensive processes, and lack of real-time capabilities. Therefore, our research team developed a method that improves performance and allows for real-time updates. In this study, we utilized an ultra-lightweight unmanned aerial vehicle (UAV) to create a non-image dataset of green vegetation. We trained a deep learning computer vision model using this dataset to recognize and segment green areas. The resulting accuracy was then computed to determine real-world coverage. Although our initial results were not as expected, mostly due to not standardize drone auto flight technique and camera intervals setting, this approach shows promise for addressing the stated problem, and further in-depth research is necessary to enhance the process and outcomes. In our study, we faced several challenges while developing and implementing our approach. One critical aspect was selecting appropriate hyperparameters for training our deep learning model. Let’s delve into some of these considerations: Mean Average Precision (mAP), Confidence Threshold, Overlap (IoU), Training Data, Error Messages during Training and Number of Epochs and Batch Size. Despite these challenges, our preliminary results demonstrate the potential of this approach. Real-time green area assessment using UAVs and deep learning holds promise for environmental monitoring. Further research will refine our methodology and address limitations.