Saline-alkali Soil Salinity Remote Sensing Inversion Model Based On BP Neural Network

Soil salinization is a major environmental issue in the world, and it is more serious in aridand semi-arid area. Acquiring accurate salinity information timely is important for monitoringand evaluating soil salinization. Traditionally, soil salinization monitoring selects fixed pointsto investigate in field, which wastes not only time but also manpower and can't showrepresentative areas. It is impossible to realize large-area, real-time inspection. Remotesensing technique shows huge excellence in these aspects. High-spectral technique is a newand effective approach in studying soil attributes. High-spectral data in ground which is thebase of band selection, validation and evaluation can build the links of ground, aviation andsatellite remote sensing data. Taking Changling County in Jilin Province as the example area, this paper aims toexplore saline-alkali soils salinity information remote sensing inversion model based on BPneural network in this semi-arid area. Firstly, the salinity of soil specimens is measured withelectric conduction method in laboratory and high-spectral data is gotten by ASD in field.Analyzing the saline-alkali soils high-spectral data characteristic, this paper discusses therelations between saline-alkali soils spectral data and salinity, then selects the best bandcombination which can represent saline-alkali soils salinity spectral characteristic by means ofmulti-variables linear stepwise regression analysis and correlation analysis methods. Exceptremote sensing factors, the degree of groundwater mineralization and buried groundwaterdepth are the two major factors which influence saline-alkali soils salinity, and these twofactors are important input variables for this model. Being an important branch of intelligence computation, neural network model is anonlinear mathematical model. This method may realize the mapping from seven dimensionsvariables to salinity information through training specimen data and adjusting the weights.Using neural network method to retrieve saline-alkali soils salinity is beneficial and can showthe potentials of geography computation techniques in analyzing high quality data. This paper was divided into four chapters as follows: The first chapter dissertates theevolvement of saline-alkali soils remote sensing inspection in the world, then express themain study goal, significance, study method and technique route in this paper. The secondchapter introduces the situations of the study area which cause soil salinization, including itsnature environment, social economic situations and the status of soil salinization. The thirdchapter quantitively analyzes the relations among the collected high-spectral data inexperiment, salinity and TM image data, and make use of multi-variables linear stepwiseregression and correlation analysis method to select the best band combination which canshow saline-alkali soils salinity characteristic. The fourth chapter designs saline-alkali soilssalinity information remote sensing inversion model based on back propagation neuralnetwork, and it includes seven input layers, that is spectral reflectance of 900nm, 800nm,760nm, 650nm , 580nm , the degree of groundwater mineralization and buried groundwaterdepth. This model contains two hidden layers, the first hidden layer including five nodes andthe second hidden layer including three nodes. It has only one output layer, that is salinityinformation. The recycling model training makes this model approach to real mappingrelation of datasets infinitely. Then this chapter gives the method of precision validation. Inthe last chapter, the results and the defects about this paper are discussed.