Hyperspectral Inversion Of Soil Total Nitrogen, Total Carbon And Carbon / Nitrogen Ratio In The Three Rivers Source Region

With the rapid development of science and technology, the increasingly maturehyperspectral remote sensing technology has become the most important and effectivemeans of estimating soil composition. The serious pollution and the continuousdeteriorating of the soil quality makes the inversion soil composition play anincreasingly important role. In this paper, we makes use of two terms data oftotal-nitrogen, full carbon, and carbon-nitrogen ratio(C/N) in the SanjiangyuanRegion and bases on spectral reflectance of soil sample chamber data measured byASD FieldSpce4to construct models by means of MSLR, PLSR, and BPNN. Theoriginal spectral reflectance (REF) was processed by nine-weighted moving averagemethod and then transforms into four forms: first-order differential (FDR), secondorder differential (SDR), the reciprocal logarithmic (Log (1/R)) and band depth (BD)modeling analysis. To highlight the differences in soil types, the overall sample andthe sample were divided into five kinds of soil types: they are alpine meadow soil,alpine grassland soil, mountain meadow soils, gray brown soil and swamp soil.Conclusions of this thesis:(1) The “Cap” of the spectral reflectance of alpine grassland soil is thecomprehensive reflection of different components of soil, which provides scientificbasis for soil type classification.(2) BPNN can stably estimate soil contents (concentration) such as total nitrogen,total carbon and nitrogen ratio in the study area. Among the results, the optimum isfull band modeling of total nitrogen (marsh soil, REF, RPD=3.47); MSLR and PLSRcannot reach such an accurate precision, but it highlights the differences between thevarious soil types and each transformation between indicators.(3) The overall sample modeling is better than the modeling of the five differentsoil type. Because the overall sample modeling is more stable and precise and most ofits five indicators’RPD are above1.4.(4) Characteristic band (500~900nm,1400~1500nm,1900~2000nm and2200~2300nm) modeling and full band (350~2500nm) modeling possess their owncharacteristic in modeling precision. Although their modeling precision is on thesimilar level, yet the Characteristic band modeling is more stable and efficient, and having less variations.(5) The inversion accuracy of C/N is less than total nitrogen and full-carbon,because during the operation about the soil C/N it can be found that the linearcorrelation between the soil spectral reflectance date and the C/N tends to weak.(6) Five kinds of transformation in the form of indicators, the transformation ofREF, Log (1/R) and BD can achieve better accuracy in three models, but FDR andSDR are more suitable in BPNN model.