Hyperspectral Soil Component Inversion Based On Kubelka-munk Model And Deep Regression Network

Soil is an essential natural resource for human production and life,which plays a crucial role in human survival and development.The cultivated land in China is facing urgent problems such as serious degradation of cultivated land quality and extensive heavy metal pollution,which poses a severe challenge to ensure food production and food security.The traditional cultivated land health monitoring ability is tightly bound by the low economic efficiency,poor performance and low timeliness.As one of the current research hotspots,the application of hyperspectral remote sensing technology can realize periodic,large-scale and high-precision cultivated land health monitoring.The research of soil composition inversion based on the hyperspectral remote sensing technology mainly relies on the traditional statistical learning method,which still has many problems,such as unknown sensitive wavebands and the lack of inversion mechanism,overfitting phenomenon under small sample data and lack of physical model-based spectral correction models for removal of soil composition interference factors.Oriented to the mechanical deficiencies of soil organic matter inversion,the theoretical of Kubelka-Munk theory is explored and introduced in this article to obtain the spectral characteristics of soil organic matter with high generalization capability.And the spectral correction model of soil moisture elimination is proposed to solve the obvious disturbance on airborne hyperspectral data caused by moisture and enhance soil organic matter's characteristics.Moreover,semi-supervised learning and deep learning methods are combined to construct the inversion model for soil organic matter,soil heavy metal As,and Cr.Finally,the spatial distribution characteristics of soil components have been analyzed.The main research contents of this article are summarized as follows:(1)To obtain the effective and sensitive spectral features of soil organic matter,an inversion model based on the thickness correction of Kubelka-Munk(K-M)theory is proposed.Firstly,a soil thickness observation experiment based on K-M theory was constructed.The impact of soil thickness and container's material on the spectra is explored by selecting different experimental samples with different material of background container.The K-M thickness model is modified by combining the indoor spectral data.After that,the corresponding scattering coefficients and absorption coefficients for soil samples with different organic matter contents are calculated.Then the scattering coefficients are linearly fitted to the soil organic matter content to obtain the sensitive band at 2.197?m.Results demonstrate the feasibility and superiority of the proposed method and further explain the sensitive bands of soil organic matter in hyperspectral data,absolute coefficient accuracy up to 0.97.Moreover,the sensitive spectral information of soil heavy metal concentration is explored.(2)To tackle the impact from soil moisture to the spectrally sensitive information of organic matter in the large-scale soil organic matter inversion mapping based on airborne hyperspectral data,a K-M-based organic matter inversion model with soil moisture influence removal is proposed to improve the characteristic expression of soil organic matter on imaging spectral data.Firstly,an unmixing method was explored to extract the spectral abundance of cultivated land in the study area.Then the inversion of soil water content in the study area was realized combining with the soil moisture inversion model based on the K-M physical mode.Finally,a moisture removal model was carried out for the spectral correction based on the quantitative description of Lambert-Beer.Experiment results show that the proposed soil moisture removal model could effectively remove the influence of soil moisture in imaging hyperspectral spectrum and highlight the spectral sensitive features towards soil organic matter,especially among the short-wave infrared range.However,the enhanced features of soil heavy metals are not obvious or even lower than its row characteristic.The results from sensitive band inversion at 0.691?m and support vector machine regression method based on the moisture-removed spectral data show that the model could effectively improve the accuracy of organic matter inversion in hyperspectral image,at least 22%improvement in accuracy of physical models and 19% improvement in accuracy of statistical regression methods.Finally,the distribution of soil colour and topographic influence were analyzed using topographic moisture index and hydrological analysisbased river network extraction methods.(3)It is difficult to extract the effective sensitive bands on the spectral feature space because of the low content of heavy metals in soil.Although it is effective for soil composition quantitative analysis,but the model accuracy of the traditional statistical methods will be affected by the limited samples.In this regard,a semisupervised deep learning regression model(Semi-DNR)is proposed to improve the accuracy of soil organic matter,soil heavy metal As and Cr content inversion under the limited samples.Firstly,a deep regression network was constructed for deep feature extraction,and a new feature combination strategy is proposed to optimize the defects of the randomness on feature selection.Then,the spatial proximity strategy guided by Tobler's First Law of Geography was introduced into the semi-supervised sample augmentation process to ensure the sample size and the validity of augmented samples and the corresponding pseudo-label,which can solve the overfitting problem in the deep regression process.Finally,pseudo-sample dynamic self-updating rule and model parameter sharing mechanism were added to improve the fine-tuning ability of the deep regression network.The results show that the Semi-DNR model has a significant superiority on accuracy.The accuracy of the test set of soil organic matter,soil heavy metals As and Cr content were 0.71,0.82 and 0.63,respectively.Finally,field investigation and the collected statistical data were combined with the spatial distribution information of soil components for pollution source analysis.The results show that the spatial distribution of soil heavy metal As is related to the adsorption and complexation of As by soil organic matter and meltwater organic matter,leading to a similar aggregation effect of As by soil organic matter.