Research On Data Mining And Integration Algorithms Of Soil Salt In Ebinur Lake Wetland On Multi-platform

Soil salinization disasters severely limit the sustainable use of land and the maintenance of ecological stability.Its strong spatial and temporal heterogeneity has caused great problems for the allocation of land resources.This problem is particularly prominent in Xinjiang.At present,remote sensing technology is the most effective method for monitoring the salinization of the soil surface,but insufficient data mining has become a bottleneck for efficient and highly accurate monitoring of salinization.Machine learning integrated algorithms can provide a powerful guarantee for data mining.In this study,the Ebinur Lake Wetland Nature Reserve was used as the research target area,and the multi-spectral data simulation of Landsat 8,Sentinel 2,and Sentinel3 was implemented using wide and narrow band simulation technology,and calculated the correlation between one-dimensional single-band and traditional index and soil EC_1:5?Using 15 kinds of transformed spectral data to construct two-and three-dimensional spectral indexes,and analyze the response relationship between the spectral index and soil EC_1:5.Using the GBRT machine learning algorithm to construct three satellites multi-dimensional soil EC_1:5 estimation models,and accurately mapping the soil EC_1:5distribution of the Ebinur Lake Reserve based on the optimal estimation model of each satellite,and provided scientific basis and reference for accurate dynamic monitoring of soil salinization and multi-platform spaceborne remote sensing evaluation.This study mainly draws the following conclusions:(1)The soil EC_1:5in the 0-10cm layer of the Ebinur Lake Wetland Reserve has a strong variability,with a coefficient of variation of 104.780%,with an average value of7.240 m S·cm^-1,and the remaining four layers of soil samples belong to medium variation;the soil in this area is alkaline,the p H distribution range of the 60-80cm soil layer is the largest,the average value is 8.902,and the coefficient of variation is 6.426%.(2)The morphological characteristics of the soil hyperspectral curve are relatively consistent.Except for the water absorption bands at 1400nm and 1900nm,and the fluctuating decline from 2200nm to 2500nm,the rest of the spectral ranges are increasing.The overall fitting effect of the three satellite multispectral simulation data is very good.The fitting curve characteristics of Sentinel 2 and Landsat 8 are very similar,but Sentinel2 fits better than Landsat 8 in the red edge band,and Sentinel 3 fits best in the visible-near infrared band.The response characteristics of different EC_1:5 to different satellite bands are more consistent.With the increase of EC_1:5,the reflectivity of different bands also increases accordingly.The three satellites have the lowest reflectance values near the blue band,and peak reflectivity and steep increase in reflectivity occur near red light and infrared light.(3)At the level of one-dimensional single band,the highest correlation coefficient of soil EC_1:5 for each of the three satellite bands is about 0.47.The sensitivity of the band near visible light to soil EC_1:5 is stronger than that of other spectral bands,and the blue band reflects the best response.In the response relationship between the traditional index and soil EC_1:5,NDVI,EVI,and NDSI failed to pass the significance test.S4 is the best salt response index with a maximum correlation of 0.4555.The response relationship of two-dimensional spectral index to soil EC_1:5 is better than that of soil salinity index as a whole,the maximum correlation coefficient is 0.7304,which is an increase of 0.2749compared with the best salt index.The correlation between the three-dimensional spectral index and soil EC_1:5 is better than the two-dimensional spectral index as a whole.The maximum correlation coefficient is 0.7657,which is 0.0353 higher than the best two-dimensional spectral index.(4)In the soil EC_1:5 estimation model based on a single variable of different dimensions,three satellites all shows the law of three-dimensional variable model>two-dimensional variable model>one-dimensional variable model.Landsat 8,Sentinel 2,and Sentinel 3 satellites have the best estimation model accuracy R² of 0.845,0.854,and 0.878,respectively.In the soil EC_1:5 estimation model based on the combination of multi-dimensional variables,the estimation accuracy of the three satellites shows a"one-dimensional+two-dimensional+three-dimensional"variable combination model>"one-dimensional+two-dimensional"variable combination model.The precision R² of each satellite's best estimation model is 0.853,0.864,and 0.887.Compared with the single-variable model,the estimation accuracy of each satellite has improved significantly,indicating that the spectral index has played a positive role in improving the model construction.For the overall analysis,Sentinel 3 is the best estimation satellite,and the"one-dimensional+two-dimensional+three-dimensional"variable combination model is the best estimation model.(5)Under the best estimation model,the overall trends of soil EC_1:5 distribution characteristics in the three satellite protection areas are relatively consistent,and the areas with high soil EC_1:5 values are mainly concentrated around the lake basin and lake area.Combined with the characteristics of each satellite load,Landsat 8 can provide strong support for inversion and long-term dynamic monitoring of soil salinization in the region.Sentinel 2 has certain application potential for fine classification and rapid dynamic monitoring of soil salinization at different scales.Sentinel 3 is more suitable for accurate monitoring and digital mapping of soil salinization on a large scale or even globally.This study combines multi-spectral satellite remote sensing with near-surface hyperspectral measurement methods,and uses machine learning algorithms to filter out the optimal variables and prediction models of soil salinization under multi-dimensional spectral parameters,providing multi-scale salinization distribution maps for different satellites.And evaluate the ability to monitor salinization in arid areas on multi-satellite platforms.It provides a scientific reference for local agricultural production management and land resource allocation,and lays a foundation for the follow-up regional-scale spatial and temporal salinization evolution trend tracking,and provides a spatial reference for the government to understand the impact of land management on the salinization process.