Utilizing residual networks for remote sensing estimation of total nitrogen concentration in Shandong offshore areas

Abstract

Nitrogen is one of the critical factors in water pollution and eutrophication, so applying the deep learning method in remote sensing inversion of nitrogen can provide basic information for environmental management. This paper proposes a two-step feature extraction method to solve the problem that the number of bands in water quality inversion is insufficient and the deep learning method cannot be fully exploited. Firstly, manual feature extraction is completed through the fusion between bands to obtain a set of high-latitude shallow factors, which make the features rich and diverse. Then, a one-dimensional convolutional residual network (ResNet-1D) is constructed, and the deep features are automatically extracted through convolution operations of the model, where the residual learning is used to reduce the training difficulty. The full connection is established through depth features. The comparison of models shows that the Mean Relative Error (MRE) is decreased by at least 10% in both test and validation datasets. Finally, the spatiotemporal distribution of total nitrogen concentration (TNC) in the coastal waters of Shandong is explored. In general, the spatial distribution is that the concentration near the coast is higher than the far. The temporal variation is that the monthly mean of the TNC is low in March, moderate in May and August, and high in October; the annual average value of TNC is 0.3mg/L, which has decreased slightly year by year since 2014.