9:30 AM - 9:45 AM
[MIS21-03] BeachLISA: Web-based System for AI Image Segmentation Analysis of Beach Litter
Keywords:Beach Litter, Image Segmentation, Web System, Marine Debris Monitoring, Quantification Method, Deep Learning
Marine debris, particularly plastic, is a critical international environmental issue. Effective countermeasures require continuous monitoring for efficient collection of beach litter. However, conventional visual observation methods face challenges such as subjectivity in evaluation, limited observation frequency, and difficulties in comparing between areas. In our previous research, we proposed a high-precision quantification method for beach litter using Deep Neural Networks (DNN). This method enables quantification of litter coverage area through image pixel classification and projective transformation. However, implementing this method required specialized knowledge in software engineering and machine learning, creating barriers to widespread adoption.
In this study, we developed "BeachLISA", Beach Litter Image Segmentation Analysis, a web system that enables automatic detection and quantification of beach litter through simple drag-and-drop image uploads in a browser, reducing these technical barriers. The system implements (1) a web browser interface, (2) web APIs capable of 128 parallel processing, and (3) slicing inference functionality for high-resolution images. In evaluation experiments, using slicing inference on high-resolution orthographic images (exceeding 10,000 pixels in width/height) captured by UAVs from above, we maintained detection accuracy for artificial litter (Precision: 68.9%, Recall: 85.1%). The parallel processing capability enabled processing of 100 images in approximately 213 seconds. Furthermore, in an application to continuous monitoring using fixed cameras, we successfully conducted observation experiments at Rokudoji Beach in Toyama Prefecture, quantitatively capturing changes in litter coverage area before and after beach cleanup events. We performed statistical analysis comparing these quantified litter coverage area results with three observational datasets: Japan Meteorological Agency observation data, the Japanese Coastal Ocean Monitoring and Forecasting System (MOVE-JPN). GPV data, and real-time wave data. These results will be discussed during the presentation.
This system enables marine pollution researchers to quantify beach litter from images without requiring specialized computer science knowledge. This is expected to facilitate observations worldwide and collection of more detailed spatiotemporal distribution data. The resulting data can contribute to solving various marine pollution research challenges, including optimization of litter collection plans, development of drift prediction models, and identification of litter sources.
In this study, we developed "BeachLISA", Beach Litter Image Segmentation Analysis, a web system that enables automatic detection and quantification of beach litter through simple drag-and-drop image uploads in a browser, reducing these technical barriers. The system implements (1) a web browser interface, (2) web APIs capable of 128 parallel processing, and (3) slicing inference functionality for high-resolution images. In evaluation experiments, using slicing inference on high-resolution orthographic images (exceeding 10,000 pixels in width/height) captured by UAVs from above, we maintained detection accuracy for artificial litter (Precision: 68.9%, Recall: 85.1%). The parallel processing capability enabled processing of 100 images in approximately 213 seconds. Furthermore, in an application to continuous monitoring using fixed cameras, we successfully conducted observation experiments at Rokudoji Beach in Toyama Prefecture, quantitatively capturing changes in litter coverage area before and after beach cleanup events. We performed statistical analysis comparing these quantified litter coverage area results with three observational datasets: Japan Meteorological Agency observation data, the Japanese Coastal Ocean Monitoring and Forecasting System (MOVE-JPN). GPV data, and real-time wave data. These results will be discussed during the presentation.
This system enables marine pollution researchers to quantify beach litter from images without requiring specialized computer science knowledge. This is expected to facilitate observations worldwide and collection of more detailed spatiotemporal distribution data. The resulting data can contribute to solving various marine pollution research challenges, including optimization of litter collection plans, development of drift prediction models, and identification of litter sources.