Domain-Knowledge-Guided Multisource Fusion Network for Small Water Bodies Mapping Using PlanetScope Multispectral and Google Earth RGB Images

Abstract

Mapping of small water bodies (SWBs) has been facilitated by very high resolution remote sensing images. The multispectral PlanetScope image, with visible to near-infrared bands at 3-m resolution, has been continuously used for mapping SWBs recently, but accurately delineating SWB boundaries remains challenging. The Google Earth image allows for mapping surface water at a finer resolution than PlanetScope, but it lacks the near-infrared band in which water and nonwater are usually distinctive. This article proposes a novel domain-knowledge-guided multisource fusion network (DKFNet), which fuses PlanetScope with Google Earth images to map SWBs at 1-m resolution while integrating the complementary information from each data. DKFNet utilizes the normalized difference water index (NDWI) image from PlanetScope image as domain knowledge for water mapping. DKFNet contains a domain-knowledge-based coordinate attention module, which has an advantage in detecting small objects, to incorporate the position and distribution information of SWBs from the NDWI image. DKFNet incorporates a domain-knowledge-based atrous spatial pyramid pooling module that extracts the multiscale features of water bodies from the NDWI image. Finally, DKFNet employs a novel reweighting loss to adjust the sample weights, enabling the network to focus on SWBs and water–nonwater boundaries, which are difficult to map accurately from traditional deep-learning networks. Results demonstrate that DKFNet predicted better water boundaries and reduced many false positives predicted by deep-learning networks using PlanetScope-only image and Google Earth-only image. DKFNet also can better map SWBs than several state-of-the-art networks using both PlanetScope and Google Earth images.

Keywords

• Deep learning
• domain knowledge
• image fusion
• sample reweighting
• small water body (SWB)