Inversion Study Of Water Quality Parameters In Huixian Karst Wetland By Combined Hyperspectral Images And Deep Learning

Karst wetlands are special and important components of karst ecosystems and play a key role in maintaining the stability of carbon sinks and improving water quality in karst areas.However,in recent years,the water environment of karst wetlands has continued to deteriorate due to the intensification of human activities.Remote sensing technology has become an important way to estimate water quality parameters because of its fast and convenient advantages,but the ability of multi-source remote sensing images to estimate water quality parameters in karst wetlands needs to be further explored.In order to solve the above problems,this paper takes the Huixian Karst Wetland in Guilin city as the study area,constructs an inversion model of water quality parameters of phycocyanin based on Sentinel-2 and ZY3 multispectral images,RF,XGB and GB algorithms,and evaluates the adaptability of shallow machine learning and satellite-based multispectral images in the inversion of water quality parameters of karst wetlands with strong optical activity.Based on deep learning（Transformer and MDN）and shallow machine learning algorithm in conjunction with multi-source remote sensing data（UAV,OHS,ZY-3and S2）,we construct inversion models of water quality parameters（Chla,PC,Turb,DO）in karst wetland,demonstrate the applicability of deep learning in the inversion of water quality parameters in karst wetlands,evaluate the difference in inversion accuracy of images with different spatial and spectral resolutions,explore the gradient distribution of concentration of water quality parameters,and conduct interpretability analysis of inversion results based on SHAP.Based on the UAV hyperspectral images and Informer deep learning algorithm and adaptive integrated learning algorithm to construct the inversion model of non-optically active water quality parameters in karst wetland,we compare the inversion accuracy of the two algorithms for total nitrogen,total phosphorus and ammonia nitrogen,compare the impact of different feature expansion methods on the estimation accuracy of water quality parameters,and conduct water quality assessment.The findings of this paper are as follows:（1）RF,XGB and GB algorithms all achieved better accuracy in the inversion of phycocyanin in karst wetlands,with R² ranging from 0.680 to 0.781;ZY3 images had higher estimation accuracy and stability than S2.（2）Transformer algorithm has the highest accuracy in the estimation of Chla,PC and DO,with R²=0.649～0.844;GB algorithm has the highest accuracy in the estimation of Turb（R²=0.752）;Transformer’s accuracy inversion for non-optically active water quality parameters is significantly higher than other algorithms.（3）The estimation accuracy of UAV images（R²=0.419～0.695）is slightly higher than that of satellite-based images,and the overall estimation accuracy of multispectral images（R²=0.338～0.718）is higher than that of OHS hyperspectral images.（4）A large area in the study area shows the trend of eutrophication of water bodies,with 73.08%of the area with Chla concentration over 40ug/L and 23.01%of the area with Turb concentration over 60 NTU,and the water pollution is more serious;Chla and DO,PC and Turb are most sensitive to NIR band and red band,respectively.（5）The accuracy（R²）of total nitrogen,total phosphorus and ammonia nitrogen estimated by the adaptive stack and Informer algorithm ranged from 0.487 to 0.631,and the estimation accuracy of Informer algorithm is higher than that of the adaptive stack;the R² is improved to the range of 0.551～0.751 by feature expansion,and the problem of overestimation and underestimation of inversion values is reduced to some extent.（6）The pollution level of total nitrogen and total phosphorus in the study area is greater than that of NH₃-N,and the concentration of total nitrogen and total phosphorus in a few areas exceeds Class IV.