Combining Active And Passive Remote Sensing And Adaptive Ensemble For Mangrove LAI Estimation Research

Accurate estimation of mangrove Leaf Area Index(LAI)and mapping of its spatial distribution is the basis of mangrove ecosystem protection and restoration.Combining active and passive remote sensing images based on machine learning(shallow machine learning and deep learning)and ensemble learning algorithm has become one of the main methods for estimating mangrove LAI.In this paper,the coastal mangrove nature reserve of the Beibu Gulf of China was taken as the research area,and images of unmanned aerial vehicles(UAV),Sentinel-2A(S2),Zhuhai-1(OHS)and Gaofen-3(GF-3)were taken as data sources to explore the mapping effect of shallow layer machine learning algorithm on the spatial distribution of mangrove LAI under multi-spectral image data.Compare and analyze the estimation abilities of shallow machine learning,ensemble learning and deep learning algorithms for mangrove LAI,and explore the impact of sample enhancement on the performance of deep learning algorithms(Transformer and DNN).The spatial distribution of mangrove LAI was mapped based on adaptive integration and deep learning algorithm combined with multi-spectral,hyperspectral and radar images,and the influence of image features on estimation of mangrove LAI was quantitatively evaluated based on SHAP.The results are as follows:(1)The shallow machine learning model(XGBoost)achieved high precision LAI estimation of different mangrove tree species,with R~2 increased by 0.105～0.365 and0.283～0.54,and RMSE decreased by 0.1～0.392 and 0.102～0.518 compared with other models.(2)The estimation accuracy of UAV image for LAI of female Marine olive was better than that of Sentinel-2A image(R~2=0.821,RMSE=0.288).Sentinel-2A images showed better LAI estimation accuracy(R~2=0.940～0.979,RMSE=0.142～0.104).(3)The optimal spectral reflectance range of optical images for mangrove LAI estimation was 650nm～680nm,and the estimation effect of GF-3 SAR images for mangrove LAI estimation with high coverage was better(R~2=0.567).(4)The integrated learning model has the best LAI estimation accuracy for different mangrove tree species(R~2=0.5266～0.713),which is 0.0019～0.149 higher than the average estimation accuracy(R~2)of deep learning models(Transformer and DNN).Sample enhancement improves the estimation accuracy of mangrove LAI in DNN and Transformer models,and 20%-40%of measured data can be reduced under the condition of achieving the same estimation accuracy.(5)The 1DCNN+DNN model realized the high-precision estimation of mangrove LAI(R~2=0.8685),and the one-dimensional convolution(1DCNN)feature extraction improved the estimation accuracy of mangrove LAI(R~2)by 0.097～0.1297 compared with the traditional DDR method.(6)GF-3 SAR and OHS hyperspectral images had the best effect in co-mapping mangrove LAI spatial distribution(R~2=0.8685,RMSE=0.134).Based on SHAP visualization analysis,it was found that image features H19[(NIR/Red)/red]、NDVI、H4[(b12-b6)/(b12+b6)]和EMVI had the largest image estimation for mangrove LAI.