Study On Hyper-Spectral Models For Predicting Black Soil Organic Matter Content

The black soil in northeast China is a kind of precious resource which isnon-renewable. It plays a very important role in grain production, economicdevelopment and the eco-environment in northeast China and even the whole country.However, the long-term human activities of over-reclamation and irrationalcultivation resulted in soil impoverishment. The water and black soil erosion is veryserious. So, it is necessary to determine soil properties in time. Protect the black soil,brook no delay. The development of hyperspectral remote sensing technologyprovides convenient and efficient means for obtaining physics and geochemicalparameter of the soil. Compared with the traditional soil investigation method, itneeds less labor, time and money. In contrast with multispectral remote sensingtechnology, its accuracy is higher. The high spectral remote sensing can providedetailed data for analyzing the surface of soil and its property. Now, it is widely usedin the inversion of the content of soil organic carbon, moisture, nitrogen, phosphorus,potassium, etc.This paper takes farmland soil in the black soil region of northeast China as theresearch object. The high-spectral data and organic matter content of68soil samplesare the main data source. Soil is a complex organism composed of many substancesand soil spectrum is the combined result of the composition. Analysis of soil spectralcharacteristics is the basis for estimating soil organic matter content. So, thedifference of laboratory soil hyperspectral reflectance under different organic mattercontent level and spectral resolution was firstly analyzed after the spectrapre-processing. Secondly, the laboratory soil spectrum was compared to the field soilspectrum. Thirdly, correlation analysis was done between black soil organic mattercontent and spectral reflectance. Then, hyperspectral predicting models of black soilorganic matter content were performed based on spectrum analysis technique andstatistics principle. The method such as stepwise multiple linear regressions (SMLR), partial least squares regressions (PLSR) and supported vector machines (SVM) wererespectively used to build models based on the black soil spectral data and itstransforms. The best predicting model was confirmed through contrasting theprecisions of different models. The main conclusions are as follows:The black soil spectral curves are gentle on the whole. Spectral reflectanceincreases with the increase of wavelength. The black soil spectral curve is controlledby organic matter. It affects the entire spectrum characteristics of the spectral curveThe difference under different organic matter level is more significant in visible bands.Without effects of the field environment, the laboratory soil spectral curve is smootherand neater. Some detail features are more obvious. The spectral response range oforganic matter in black soil is wide. The curve became smoother with little changewhen it was resampled to10nm. However, the soil spectral absorption characteristicswere gradually weakened when the resampling interval was up to20nm. Higherorganic matter content is associated with the lower spectral reflectance.550~680nmcan be used as the diagnosis wave bands of the organic matter in black soil.Correlation between organic matter and differential coefficient of absorb reflectance([lg(1/R)]) is much more significant than that between organic matter and othertransforms. The maximum coefficient of correlation is0.80at1274nm.Compared with SMLR predicting models established by original reflectancespectra and other forms of data transformations, the [lg(1/R)] and CR data conversionmethods were better. The precisions of all predicting models were improved with thedependent variable of lg(SOM). The precisions of predicting models based on partialleast squares regression (PLSR) method were better than SMLR method. The value ofRPD was2.664. So, the model can be used for external validation. The precisions ofPLSR models increased and decreased with spectral resolution, and the maximumaccuracy is at10nm. It indicates that the influence of noise was weakened afterresampling. The precisions of nonlinear models based on least squares support vectormachine regression (LS-SVR) method were better than the former two linear models.In addition, the LS-SVM models with feature extraction using PLS had higherprecision and better generalization ability. The value of R2Cwas0.993and the RPDreached3.682. The gap between the modeling precision and predicting precision hadbeen narrowed.