Scientists Develop Novel Method to Detect Polluted Water in Rural Areas, Tackling Algal Blooms
February 28, 2025
Chinese scientists have developed a novel method to detect and monitor polluted water in rural areas—a critical step toward tackling harmful algal blooms and contaminated ponds. By combining high-resolution satellite imagery with deep learning methods, researchers from the Aerospace Information Research Institute (AIR) under the Chinese Academy of Sciences (CAS) can now pinpoint ponds overrun by duckweed or algae with remarkable accuracy.
Published in the Journal of Cleaner Production, the study offers a powerful new tool for identifying nutrient-rich or heavily polluted water bodies, parts of which often emit foul odors and threaten ecosystems. This innovation could transform how communities track and address water quality issues, paving the way for healthier environments in vulnerable regions.
With the increase of nutrients in water, algae and duckweed proliferate extensively in ponds and ditches. Water seasonally or annually covered by duckweed or algal blooms are defined as Duckweed and Algal Bloom Water (DAWs) in this study. DAWs in basins with high organic matter content or eutrophication are particularly prone to becoming black and odorous. Therefore, identifying DAWs to assist in screening black and odorous water, which plays a significant role in environmental monitoring and management.
The research team developed a lightweight model using high-resolution remote sensing imagery (0.25m) and deep learning methods. To improve the model's ability to identify DAWs, they introduced three key indicators—chromaticity angle (α), slope sum (S), and green channel index (GI). These indicators helped create a new band combination (ASGI), which made the DAWs more distinguishable in remote sensing images and boost the model's accuracy.
The team also added a special attention module (scSE) to the lightweight MobilenetV2-Unet architecture, improving its ability to detect DAWs in complex backgrounds. The model SEM-Unet, remain efficient, but can now accurately identify DAWs even in challenging, feature-rich backgrounds.
The Haihe River Basin, a region grappling with severe eutrophication, served as a critical testbed for the study, which not only accurately identified DAWs but also pinpointed rural black and odorous water bodies. The success shows that the approach can provide efficient technical support for water environment monitoring.
The research combines high-resolution imagery with deep learning methods to develop a lightweight, high-precision algorithm for identifying DAWs in complex environments. It also provides an efficient method for rapidly screening large-scale rural black-odor water bodies in basins with high organic matter content or eutrophication.
Comparison of prediction results between the model inputs using RGB and ASGI methods. It is evident that the RGB input method exhibits significant false positives for easily confusable features, such as red rooftops. (Image by AIR)
Comparison of recognition results of different semantic segmentation networks. (Yellow boxes highlight areas with over-predicted boundaries, red boxes highlight areas with under-predicted boundaries) (Image by AIR)
Research News
Scientists Develop Novel Method to Detect Polluted Water in Rural Areas, Tackling Algal Blooms
Chinese scientists have developed a novel method to detect and monitor polluted water in rural areas—a critical step toward tackling harmful algal blooms and contaminated ponds. By combining high-resolution satellite imagery with deep learning methods, researchers from the Aerospace Information Research Institute (AIR) under the Chinese Academy of Sciences (CAS) can now pinpoint ponds overrun by duckweed or algae with remarkable accuracy.
Published in the Journal of Cleaner Production, the study offers a powerful new tool for identifying nutrient-rich or heavily polluted water bodies, parts of which often emit foul odors and threaten ecosystems. This innovation could transform how communities track and address water quality issues, paving the way for healthier environments in vulnerable regions.
With the increase of nutrients in water, algae and duckweed proliferate extensively in ponds and ditches. Water seasonally or annually covered by duckweed or algal blooms are defined as Duckweed and Algal Bloom Water (DAWs) in this study. DAWs in basins with high organic matter content or eutrophication are particularly prone to becoming black and odorous. Therefore, identifying DAWs to assist in screening black and odorous water, which plays a significant role in environmental monitoring and management.
The research team developed a lightweight model using high-resolution remote sensing imagery (0.25m) and deep learning methods. To improve the model's ability to identify DAWs, they introduced three key indicators—chromaticity angle (α), slope sum (S), and green channel index (GI). These indicators helped create a new band combination (ASGI), which made the DAWs more distinguishable in remote sensing images and boost the model's accuracy.
The team also added a special attention module (scSE) to the lightweight MobilenetV2-Unet architecture, improving its ability to detect DAWs in complex backgrounds. The model SEM-Unet, remain efficient, but can now accurately identify DAWs even in challenging, feature-rich backgrounds.
The Haihe River Basin, a region grappling with severe eutrophication, served as a critical testbed for the study, which not only accurately identified DAWs but also pinpointed rural black and odorous water bodies. The success shows that the approach can provide efficient technical support for water environment monitoring.
The research combines high-resolution imagery with deep learning methods to develop a lightweight, high-precision algorithm for identifying DAWs in complex environments. It also provides an efficient method for rapidly screening large-scale rural black-odor water bodies in basins with high organic matter content or eutrophication.
Comparison of prediction results between the model inputs using RGB and ASGI methods. It is evident that the RGB input method exhibits significant false positives for easily confusable features, such as red rooftops. (Image by AIR)
Comparison of recognition results of different semantic segmentation networks. (Yellow boxes highlight areas with over-predicted boundaries, red boxes highlight areas with under-predicted boundaries) (Image by AIR)