Lightweight Deep Neural Network Method for Water Body Extraction from High-Resolution Remote Sensing Images with Multisensors.

Abstract

Rapid and accurate extraction of water bodies from high-spatial-resolution remote sensing images is of great value for water resource management, water quality monitoring and natural disaster emergency response. For traditional water body extraction methods, it is difficult to select image texture and features, the shadows of buildings and other ground objects are in the same spectrum as water bodies, the existing deep convolutional neural network is difficult to train, the consumption of computing resources is large, and the methods cannot meet real-time requirements. In this paper, a water body extraction method based on lightweight MobileNetV2 is proposed and applied to multisensor high-resolution remote sensing images, such as GF-2, WorldView-2 and UAV orthoimages. This method was validated in two typical complex geographical scenes: water bodies for farmland irrigation, which have a broken shape and long and narrow area and are surrounded by many buildings in towns and villages; and water bodies in mountainous areas, which have undulating topography, vegetation coverage and mountain shadows all over. The results were compared with those of the support vector machine, random forest and U-Net models and also verified by generalization tests and the influence of spatial resolution changes. First, the results show that the F1-score and Kappa coefficients of the MobileNetV2 model extracting water bodies from three different high-resolution images were 0.75 and 0.72 for GF-2, 0.86 and 0.85 for Worldview-2 and 0.98 and 0.98 for UAV, respectively, which are higher than those of traditional machine learning models and U-Net. Second, the training time, number of parameters and calculation amount of the MobileNetV2 model were much lower than those of the U-Net model, which greatly improves the water body extraction efficiency. Third, in other more complex surface areas, the MobileNetV2 model still maintained relatively high accuracy of water body extraction. Finally, we tested the effects of multisensor models and found that training with lower and higher spatial resolution images combined can be beneficial, but that using just lower resolution imagery is ineffective. This study provides a reference for the efficient automation of water body classification and extraction under complex geographical environment conditions and can be extended to water resource investigation, management and planning.