Effects of Spectral Resolution and Spectral Preprocessing on the Estimation Accuracy of Soil Organic Matter Content

Abstract

The use of hyperspectral remote sensing data combined with statistical modeling is currently the main method for hyperspectral estimation of soil organic matter (SOM) content. Hyperspectral remote sensing has high spectral resolution and large data volume. Generally, data redundancy is reduced by dimensionality reduction and de noising, but improper processing may affect the accuracy of SOM estimation. In order to explore the influence of spectral resolution and spectral preprocessing on SOM content estimation and seek the optimal spectral resolution for SOM content estimation, the study takes soil samples from the black soil area of Heilongjiang Province as the research object, uses the SR-6500 portable spectrometer to measure the spectral reflectance of the soil sample in the laboratory, and resamples the spectrum to 12 kinds of spectral resolution such as 1nm, 5nm, 10nm, 20nm, 30nm, 40nm, 50nm, 60nm, 70nm, 80nm, 90nm and 100nm. We use the Savitzky-Golay smoothing method to perform spectral smoothing and denoising, and perform 9 kinds of spectral mathematical transformation preprocessing such as reflectance reciprocal(1/R), logarithm(LnR), reciprocal logarithm(Ln1/R), logarithm reciprocal (1/LnR), first-order differential (R‘), reciprocal first-order derivative ((1/R)‘), logarithmic first-order derivative ((LnR)‘), reciprocal logarithm first-order derivative ((Ln1/R)’) and logarithmic first-order derivative((1/LnR)‘). Finally, we use multiple linear regression (MLR) and partial least square regression (PLSR) methods to establish the SOM content estimation model. The results show: (1) The spectral resolution of MLR model SOM estimation and verification accuracy determination coefficient R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> val >0.94 is: 70nm > 5nm> 10nm. And the best resolution is 70nm preprocessed by 1/R, R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> val =0.958, RMSEP=1.516; (2) The spectral resolution of PLSR model SOM estimation verification accuracy R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> val >0.94 is: 20nm>40nm>70nm>10nm>5nm. And the best resolution is 20nm preprocessed by (1/R)‘, R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> val =0.953, RMSEP=1.606, RPD=4.24; (3) Comparing the estimation accuracy of SOM content with different reflectance mathematical transformations, we find that 1/R and (1/R)’ preprocessing have a better effect on improving the estimation accuracy. Among them, the model verification accuracy R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> val of 1/R preprocessing is higher than 0.87, and the model verification accuracy R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> val of (1/R)’ preprocessing is higher than 0.90; (4) SOM content models with 5nm, 10nm, 20nm, 40nm, 70nm spectral resolution have high and stable accuracy. Therefore, the spectral resolution used for SOM content estimation is not as high as possible. Appropriate reduction of spectral resolution and selection of appropriate spectral preprocessing methods can not only reduce the workload of data processing, but also improve the estimation accuracy of SOM content.