Rapid Acquisition Of Soil Information And Management Zones Delineation Based On Multi-source Data Fusion At Field Scale

Precision agriculture is the main trend of agricultural development in China,and their development is closely related to ecology,environment,and economy.Soil testing and fertilizer recommendation is one of the main contents for precision agriculture,and it is of great significance to reduce resource waste,protect soil and environment,and improve crop yield.Fast and accurate measurement of soil information,and then delineating farmland management zones scientifically and reasonably is the foundation of fertilizer recommendation.Traditionally,soil information measurement was based on complicated soil sampling and chemical analysis,which was inefficient and caused problems such as environment pollution,soil destruction,etc.With the development of proximal soil sensing and remote sensing,an increasingly number of sensors are applied in soil information acquisition.However,it is difficult for robust and thorough measurement of soils over large area using single sensor duce to the complexity and comprehensiveness of soil.Therefore,this paper focused on multi-source information fusion,including proximal soil sensor fusion,remote sensing and proximal sensing fusion,multi-temporal data fusion by using various sensors and satellites,such as vis-NIR,MIR,EM38,TDR,GF1,Sentinel-1,Sentinel-2,MODIS,etc.for soil attributes fast evaluation,crop growing analysis and farmland management zones delineation.The main contents and results were as follows:（1）Soil properties estimation based on proximal soil sensor fusionSoil properties estimation using single sensor was usually less stable,so we tried sensor fusion for soil estimation to improve the accuracy and stability of the estimation.Six soil properties including SOM,p H,TN,AN,AP and AK were estimated using visible near infrared diffuse reflectance spectroscopy（vis-NIR）and mid infrared diffuse reflectance spectroscopy（MIR）independently as well as three levels fusion of these two sensors.Sensor fusion were realized by using Boruta feature selection,outer product analysis（OPA）and Granger–Ramanathan averaging（GRA）.And we divided sensor fusion into three levels:low-level fusion was achieved by the direct combination of vis-NIR and MIR spectra;middle-level fusion was achieved by the combination of featured bands of vis-NIR and MIR spectra using Boruta algorithm（Boruta-vis-NIR&Boruta-MIR-PLSR）;high-level fusion included OPA fusion（Boruta-vis-NIR?Boruta-MIR-PLSR）and GRA model fusion.The results indicated that the estimation accuracy based on MIR spectra for six soil properties were better than that of vis-NIR spectra.Compared single sensor estimation model with three-level fusion model,middle-and high-level fusion model got better estimation results than MIR.Among these models,OPA fusion got the best results for six properties from both accuracy and model uncertainties.Among six properties,SOM,p H,TN and AN could be predicted successfully with₍（8） larger than 0.80,and RPIQ larger than 2.0.AK could also be estimated well with₍（8） of 0.77,RPIQ of 2.0.And for AP,OPA fusion could achieve acceptable accuracy with₍（8） of 0.60,RPIQ of 1.70.These results confirm the significant improvements and effectiveness of OPA fusion based on featured variables for soil properties estimation and it can provide reference for the comprehensive application of proximal soil sensors.（2）Spatial estimation and mapping of six soil properties based on remote sensing and proximal sensing fusionvis-NIR spectra could predicted soil properties such as SOM,TN etc.successfully,as proved by previous researches.While vis-NIR spectra measurement was still based on points and it is difficult to realize spatial estimation over large area.To solve this problem,this study fused remote sensing data with proximal sensing data to make full use of the advantages of high accuracy of vis-NIR spectra and wide coverage of remote sensing.Spatial estimation and mapping of six soil properties were achieved by integrating remote sensing and proximal sensing data and using three different models including partial least squares regression（PLSR）,machine learning（RF）and deep learning（CNN）.The results indicated that the estimation accuracy of each soil property was improved largely by fusing remote sensing data and proximal sensing data.The relative improvement in%RMSE for SOM,p H,TN,AN,AP and AK were 10.61%,20.00%,28.57%,15.29%,18.13%and 15.09%,respectively.Compared three different models we found that the estimation accuracy for six properties using CNN model was higher than that of PLSR or RF while its model uncertainties was lower than others,thus,CNN model produced best estimation for each property.In conclusion,p H and SOM could be predicted well using remote sensing data and proximal sensing data fusion with CNN model(_{（8）>0.70,RPIQ>2.0};TN and AN could be predicted properly with₍（8） larger than 0.64 and RPIQ larger than 1.60.Compared with traditional interpolation results,method in this study produced similar spatial distribution with more detailed information.Thus,it could provide reference for soil properties mapping at farmland scale.（3）Farmland management zones delineation based on multi-source information fusion and multi-temporal data fusionThe indicators of farmland management zoning are diverse and complex,and the single temporal remote sensing data is difficult to represent the stable growth state of crops.To solve this problem,this study proposes a multi-source data and multi temporal data fusion for farmland management zone delineation.We made a comprehensive use of multi-source and multi-temporal data including apparent soil electrical conductivity（ECa）data from EM38,soil water from TDR and Sentinel-1 and Sentinel-2 data,NDVI from multi-temporal GF 1 data and soil properties form CNN model in the second part.Besides we integrated geostatistics,deep learning（MLP）and time series analysis to construct index system of farmland management division which can fully characterize the spatial and temporal heterogeneity of soil and crop.Based on that,robust geographical weighted principal component analysis（RGWPCA）and possibilistic fuzz c-means clustering（PFCM）were used for management zones delineation.The results indicated that:soil moisture inversion model based on radar backscattering coefficient from Sentinel-1 and spectra index from Sentinel-2 using MLP model was good with₍（8）of 0.77 and RPIQ of 1.87,and its spatial distribution trend was highly consistent with the traditional interpolation results.Using MODIS 8-day time series data to extract the wheat growth lines in the study area from 2016 to 2020,it was found that the wheat growth lines in 2016-2019 were similar,and the two key phenological periods:turning green period and heading stage of wheat were in the 33rd and 116th days respectively.Compared with other years,the growth of wheat in 2020 is obviously better.Due to the better soil moisture and temperature conditions in the early growth stage,the growth of wheat is in a rising state before heading stage,and the turning green stage is not obvious,while the heading period is similar with other years.The optimal number of zones based on RGWPCA and PFCM was three and properties in each zone were significant different which indicated the zoning results were reasonable and could provide basis for precise fertilization and farmland management.