Abstract
This study investigated Baiyangdian Lake in the Xiong’an New Area to invert total phosphorus
(TP) concentrations using Sentinel-2 multispectral imagery and in situ measurements, aiming to
support water-quality assessment and eutrophication management. Based on the multispectral remote
sensing data of Sentinel-2 and measured total phosphorus data of water bodies, the study used specific
feature combinations such as NDTI, B5, B11-B12 difference, and sensitive combination bands of
measured data as model inputs to construct a PSO-Adam-BP neural network machine learning model
for total phosphorus concentration inversion. Compared with the traditional BP and PSO-BP baseline
models, the proposed framework significantly improved the accuracy of the model, with an R2 value
reaching 0.9351. Moreover, it successfully reduced the average relative error by up to 53.5%, from
4.80% to 2.24%, and decreased the maximum relative error by up to 31.7%, from 11.21% to 7.66%,
demonstrating the model’s highly robust and precise inversion performance. This study provides a new
method for water quality detection in Baiyangdian Lake and is of great significance to the protection
and development of the water environment in Xiong’an New Area.