Investigation On Remote Sensing Algorithm Of Chromophoric Dissolved Organic Matter (CDOM) In The Pearl River Estuary With Machine Learning Methodology

Chromophoric dissolved organic matter（CDOM）plays a crucial role in aquatic ecosystems,significantly impacting the carbon cycle and biogeochemical processes.The use of satellite image data for the retrieval of aquatic CDOM has become an es-sential means of studying the effects of CDOM.In this study,based on the CDOM ab-sorption coefficients and in-situ spectral remote sensing reflectance of the Pearl River Estuary（PRE）,six machine learning algorithms were developed to construct CDOM in-version models.The performance differences among the machine learning algorithms were compared,and the internal structure and principles of the models were explored in-depth to better understand their capabilities in CDOM retrieval.Additionally,the impact of tides and wind on the distribution of CDOM was studied,revealing the spa-tiotemporal variation characteristics of CDOM in the PRE.The main findings of this study are as follows:（1）Six machine learning models were trained to develop CDOM algorithms,in-cluding Support Vector Regression（SVR）,Random Forest（RF）,Extreme Gradient Boosting（XGBoost）,Multilayer Perceptron（MLP）,k-Nearest Neighbors（KNN）,and Convolutional Neural Network（CNN）.The results showed that among the machine learning models,the XGBoost algorithm performed best,with the highest²value（0.90）and the lowest CDOM root mean square error(0.37 m^-1).Compared with empir-ical algorithms,these machine learning algorithms had²values above 0.78,with the Extreme Gradient Boosting（XGBoost）performing particularly well.（2）The XGBoost algorithm identified B4/B1 as the most critical input parameter,with a contribution of71%,followed by B3/B2 with a 16%contribution.B1,B2,B3,and B4 are the wave-length bands of the OLI sensor on the Landsat satellite.These two band-ratios accounted for most of the contributions,suggesting their significant role in CDOM retrieval from Landsat OLI images.（3）Further analysis of the optimized models using the SHAP（SHapley Additive ex Planations）interpretable model revealed that band combinations played a crucial role in CDOM estimation,especially the three-band index”TBI”con-tribution to the model（average shap value of 0.82）.It was also found that the band ratio”BR1”introduced some uncertainty in the XGBoost model,while the RF model showed a more balanced performance.（4）Using Landsat-8 satellite images,we inverted six in-stances of CDOM spatial distribution patterns in the PRE.These instances revealed that CDOM in the PRE was influenced by various factors.In the PRE region,tides and wind are the primary drivers of spatial and temporal variations of CDOM.Therefore,for the management and protection of aquatic ecosystems in the PRE and its surrounding areas,it is essential to use highly efficient machine learning methods to monitor the trends of CDOM changes.This study provides a reference for water quality management and coastal environmental protection in the PRE.