Research On Vegetation Water Content Retrieval Algorithm For Colligating Machine Learning Models And Multi-Source Data

Vegetation water content(VWC)is considered as a key variable to evaluate vegetation physiological status.In agriculture,the study of VWC can guide accurate irrigation and predict yield.In ecological research,the change of VWC is of great significance to the analysis of environmental changes,the study of abnormal climate phenomena and natural disasters.Therefore,it is important to obtain VWC products with high precision and long time series.The traditional method of VWC measurement is field sampling,drying and weighing,which is destructive,time-consuming and laborious,and it is impossible to obtain historical data.The Global Navigation Satellite System Interferometric Reflectometry(GNSS-IR)technique has provided us with an effective approach for the monitoring of VWC.The normalized microwave reflection index(NMRI)was defined to reflect the change of VWC,based on the fact that the amplitude of the direct and reflected GNSS interferometric signal is related to the variation of VWC.However,the sparse distribution of the ground-site-based observation stations restricts the application of NMRI.Fortunately,remote sensing technology has the advantages of wide coverage and long observation time,which can make up for the shortage of site-based data.Therefore,in this study,the spatiotemporally continuous VWC retrieval algorithm was proposed based on the machine learning models and point-surface fusion technique by fusing GNSS-IR,optical and microwave remote sensing VWC-related datasets to miner VWC information in multisource data,and gather the advantages of multi-source data.Key work and innovations are shown below:(1)Based on the Random forest(RF)model,VWC retrieval algorithm by fusing optical remote sensing vegetation parameters and GNSS-IR NMRI dataset was proposed to realize the retrieval of VWC products with high spatial resolution.Firstly,the optimal auxiliary optical remote sensing datasets(including GPP and NDVI)was determined by correlation analysis.Based on the principle of point-surface fusion,the retrieval model was trained,with the auxiliary datasets used as the model input and the NMRI dataset as the model output.The retrieval accuracy of retrieval models was validated by ten-folder cross validation method.The results show that the machine learning model is obviously superior to the linear model,and the RF model performed the best.Then,using the RF model,spatially continuous VWC products with 500 m spatial resolution are obtained,and the feasibility of using VWC products with more fine resolution to predict and analyze drought events is verified by using the long time series retrieval results.(2)Based on GRNN model,VWC retrieval algorithm by fusing microwave remote sensing vegetation parameters vegetation optical depth(VOD)and GNSS-IR NMRI dataset was proposed to successfully extracted VWC information from VOD datasets and realize the retrieval of VWC products with temporal spatial resolution and spatial continuity.Based on the point-surface fusion principle,the retrieval model was trained,with the VOD datasets of multiple bands used as the model input and the NMRI dataset as the model output.The results show that the GRNN model has the highest accuracy compared with other traditional models.Using the GRNN model,we obtained the VWC retrievals with 25 km spatial resolution and one day temporal resolution.By analyzing the long time series variation diagram and the annual average deviation distribution map of the retrievals and precipitation,the response of VWC to the abnormal climate phenomenon of El Ni(?)o-Southern Oscillation(ENSO)was obtained.The results showed that NMRI of different vegetation types responded differently to ENSO.Compared with precipitation,the abnormal performance of VWC has lag effect.