Abstract: Black and odorous water bodies are widely distributed and seriously damages people’s living environment and the overall beautiful image of the city. With Langfang City in Hebei Province as the research area, we conducted a comparative experimental study on the extraction of remote sensing information of black and odorous water bodies by using PSPNet (Pyramid Scene Parsing Network) and U-Net model based on the multi spectral data and measured data of GF-2 satellite. The normalized difference vegetation index (NDVI) and normalized difference black odorous water index (NDBWI) are calculated based on the visible light band (RGB) and near infrared band (NIR). In view of the common problem of missing detection of small black and odorous water bodies, the module of attention mechanism is introduced to optimize and improve the model. Thus, an improved deep learning model of extracting the remotely-sensed information of black and odorous water bodies is constructed. Results demonstrate that the U-Net model, which inputs RGB+NIR+NDVI+NDBWI six channel remote sensing images in combination with attention mechanism, has the most superior extraction result for black and odorous water bodies, and achieves 0.864 5, 0.868 1 and 0.835 9 in accuracy evaluation indexes F1-srore, MIoU and Recall, respectively.

Key words: black and odorous water, deep learning model, PSPNet model, U-Net network, GF-2 satellite, remotely-sensed information, attention mechanism