| Soil salinity is a highly random land degradation problem with a complex formation mechanism,which is susceptible to climate change,water cycle and human activities.Globally,nearly one-fifth of agricultural and irrigated land is suffering from salinity,especially in arid areas,where salinity has become one of the core problems limiting sustainable agricultural development,affecting human production and life,and threatening ecological security in dry areas.Traditional methods based on field experiments and numerical simulations cannot characterize the occurrence of salinization changes at regional scale due to long observation period and small coverage,and it is also difficult to meet the demand of monitoring soil salinization dynamics in a large scale and long time series.With the development of spatial information science,multi-source satellite data and multi-modal remote sensing techniques have been widely used for salinization monitoring due to their advantages of macroscopic,short acquisition period and strong process expression capability.However,due to the differences in spatial,temporal and radiometric resolutions of different satellite data,the multi-grain size characteristics of satellite data and geo-cognitive processes have not been fully integrated in current research,resulting in deficiencies in local salinization characterization.Therefore,on the one hand,there is still a need to explore the applicability of different satellites for soil salinity monitoring,and on the other hand,the modeling process should consider the application potential of models combining different environmental variables(climate,topography,soil parent material,indices,etc.)based on soil genesis.More importantly,due to the high spatial heterogeneity of soil salinity and the scale effect problem of remote sensing monitoring,the use of machine language learning algorithms in artificial intelligence combined with knowledge of soil genesis to model soil salinity in terms of geotemporal variability in order to fully reveal its occurrence,evolution and driving mechanisms is a frontier area of research in the coming period.The Ebinur Lake Basin is an important core area of "One Belt,One Road",and salinization is a constraint to its ecological environment,food security and water resources utilization.Based on this,this paper takes the Ebinur Lake Basin as the target area,firstly,we use multi-source remote sensing satellite data(Sentinel-1A,Sentinel-3A,Landsat-8 data)and various environmental variables(DEM,meteorological data,etc.),and explore the different modeling strategies in combination with machine learning algorithms according to the pure remote sensing model,pure environmental variables model and integrated remote sensing + environmental variables model.The optimal model for remote sensing monitoring of soil salinity in the Ebinur Lake Basin under different combinations of modeling strategies is explored.Secondly,based on the spatial and temporal fusion of multi-source satellite data,the effect of salinity simulation based on the optimal model at different spatial resolution scales is evaluated for multimodal remote sensing and its scale effects.Finally,based on the optimal model construction and scale effect assessment,the multi-year soil salinity change pattern in the Ebinur Lake Basin was clarified,and the driving factors of soil salinity change in the basin were analyzed using land use change and climate data.The study provides a reference solution for rapid,effective and large-scale monitoring of soil salinity and research on the occurrence mechanism of soil salinity in arid areas.The main results and conclusions of this paper are as follows.(1)Eight modeling strategies were obtained by multimodal remote sensing combined with artificial intelligence machine learning models,and the eight strategies were modeled by four machine learning models(XGBoost,RF,SVM and PLSR).Comparing the accuracy of different modeling strategies and machine learning algorithms,the results show that XGBoost model has the best modeling accuracy for this study area,and Landsat-8 multispectral band data modeling combined with XGBoost has the best modeling inversion accuracy and good mapping effect.Based on this,the best modeling results and the best applicability were obtained by using the original image band data for remote sensing salinity monitoring work in the Ebinur Lake Basin.And the machine learning model for multi-grain soil salinity inversion can compensate for the deficiencies of remote sensing for local representation.(2)The spatial resolution is divided into scales of 30 m and modeled at seven different scales.The optimal resolution selection and image fusion algorithms are used to construct a soil salinity model for a typical inland closed watershed and verify its feasibility,providing a new solution for higher quality soil salinity prediction mapping at the watershed scale in arid zones.The estimated soil salinity increases with the scale and satisfies the nonlinear relationship of Gaussian function.The modeling effect varies with increasing spatial resolution,and the image information is gradually blurred over a wide range.Three data fusion methods,STARFM,ESTARFM and STDFA,were selected to fuse high-resolution Landsat-8 data and low-resolution MODIS data and participate in soil salinity modeling,and it was found that the STARFM method was more applicable to the Ebinur Lake Basin,indicating that the fused images could solve the problem of missing images caused by sampling time difference or cloudiness.(3)The characteristics of soil salinity distribution in the Ebinur Lake Basin in the last decade are more stable and show strong spatial heterogeneity,which provides a new research direction for quantitative soil salinity monitoring under watershed scale and long time span.The salinity data were analyzed by spatial analysis methods(Moran’s I and LISA),and it was found that soil salinity has an obvious clustering effect in spatial and temporal distribution,i.e.,the soil salinity content changes similarly within the same landscape,and the property characteristics are similar in the same soil-forming environment.The analysis of the change in the slope of the time series revealed a decreasing trend of salinity in the wetlands around the lake,and a fluctuating change in salinity in the farmland area inside the oasis and the desert area east of the lake.(4)The analysis of drivers of soil salinity changes in the Ebinur Lake Basin is based on meteorological data(temperature,precipitation,and evapotranspiration)and land use types.The drivers of soil salinity change in the Ebinur Lake Basin were analyzed mainly with meteorological data(temperature,precipitation and evapotranspiration)reflecting climate change and land use data reflecting human activities.Comparing the changes in drought indices at six time scales(January,March,June,September,December and 24 months)for four meteorological stations in the Ebinur Lake Basin,it was found that the frequency of drought increased with increasing time scales.The temperature showed an increasing trend in the northwestern and northeastern parts of the basin,precipitation showed a decreasing trend in the northwestern and northeastern parts of the basin,and the larger basin evapotranspiration was mainly concentrated in the eastern and western parts of the basin.soil salinity showed fluctuating changes in the types of cropland to cropland and grassland to cropland during 2010-2020,and soil salinity showed a decreasing trend in the types of desert to desert,grassland to grassland and desert to grassland.Taken together,the degree of soil salinity is exacerbated by climatic drought,and the positive or negative effects of human activities on soil salinity are influenced by the change of land use types. |
PDF Full Text Request