Remote Sensing Inversion Of Soil Organic Matter In Cultivated Land At Field Scale

Land is the most important carrier of human production and life,agricultural production is the oldest and most main form of land use.The Northeast Black Soil Region is one of the most important grain production bases in our country,which bears the important task of national grain strategic security.However,the extensive management,excessive application of fertilizer and pesticide cause soil consolidation,farmland degradation,agricultural productivity is difficult to continue to improve,and it is difficult to guarantee the national grain strategic security.Evaluation of cultivated land quality and precise fertilization of cultivated land is the only way to solve the problem of cultivated land and the healthy and sustainable development of agriculture.Soil Organic Matter content is the most important representative of soil nutrients,and is also one of the main indicators of cultivated land quality evaluation.Obtaining the spatial difference distribution of SOM in the surface layer of cultivated land in time is helpful to evaluate the quality of cultivated land and carry out variable fertilization.Traditional methods for investigating soil organic matter are time-consuming,laborious and costly.The rapid development and effective application of remote sensing technology provide a new approach and favorable conditions for SOM prediction.In this paper,41.3 hectares of fields in Helun City in the black soil region of Heilongjiang Province as the study area.Two methods of establishing linear regression prediction model in different subregions and establishing BP neural network prediction model are used to improve the prediction accuracy of SOM.The BP neural network prediction model method is verified by using30 hectares of field in Zhaoguang Farm,and the feasibility of this method is proved.The data of two periods of remote sensing images(Landsat 8 on May 17,2016 and Sentinel-2A on May 25,2016)in the bare soil period and 4-meter spatial resolution digital elevation model(DEM)of Helun City field were acquired,and the data of two periods of remote sensing images(Sentinel-2A on April 29,2017 and Landsat 8 on May 11,2017)in the bare soil period and 5-meter spatial resolution DEM of Zhaoguang Farm field were acquired.The relationship between image spectral information,topographic factors and soil organic matter was analyzed.The spatial variation law of soil organic matter and chemical properties,reflectance spectrum information and topographic factors are cleared.Based on image and DEM data,Helen City field was subregion by object-oriented Multi-scale segmentation and Hierarchical clustering.Then SOM linear regression prediction models for whole study area and every sub-regions were constructed.At Helun City field and Zhaoguang Farm field,the relationship between single phase image and SOM wasanalyzed,changes of soil water content in two-phase images and topographic factors on the accuracy of SOM prediction model was analyzed,the SOM prediction model based on BP(back propagation)neural network was established.The two methods in this paper can effectively improve the precision of SOM inversion and accurately monitor the spatial distribution of SOM in cultivated land.The establishment of linear regression prediction model in different regions shows that topographical affects the spatial distribution of SOM,and the accuracy of linear regression model for the whole block is low.This provides a theoretical support for the research of SOM prediction in this paper,which establishes BP neural network prediction model and adds topographic factors as input.The results show that:(1)considering the field as a whole,the linear regression prediction model of SOM is less accurate(R2 = 0.16,RMSEcal = 1.61,RMSEval = 1.45).(2)By dividing the whole field into four sub-regions: "sedimentary area","sedimentation-buffer area","erosion-buffer area" and "erosion area",and considering the SOM content of each subregions and their spectra,topographic characteristics to build SOM prediction model of each region.The total accuracy of SOM prediction in the field is obviously improved(R2= 0.58,RMSEcal = 1.17,RMSEval = 1.30).(3)The establishment of BP neural network prediction model for the whole field can effectively explore the nonlinear relationship between the spectrum,topography and SOM,and significantly improve the SOM inversion accuracy of the field.(4)For the entire field,using single-phase image,two-phase image modeling,two-phase image and topography factor to establish SOM prediction model of BP neural network,respectively.The accuracy of the model has been improved significantly,and the precision of BP neural network based on two-phase image and topography factor is the highest(R2cal = 0.917,RMSEcal = 0.492;R2val = 0.928,RMSEval = 0.0.499).(5)The SOM inversion of Zhaoguang Farm Block was successful(R2cal = 0.905,RMSEcal =0.197;R2val = 0.896,RMSEval = 0.239).It is further proved that the SOM prediction model of BP neural network based on remote sensing image,terrain factor,time phase information and terrain factor can effectively improve its accuracy,and it is an effective SOM inversion method.The results of this study can provide reference for the study of SOM inversion at field scale,and provide technical support for the evaluation of cultivated land quality and precision fertilization of cultivated land.It is of great significance in improving agricultural production efficiency,protecting cultivated land quality,and providing cultivated land information for government agricultural land management.