Predicting field capacity, wilting point, and the other physical properties of soils using hyperspectral reflectance spectroscopy: two different statistical approaches

• Published in: Environmental monitoring and assessment Vol. 186; no. 8; pp. 5077 - 5088
• Main Authors: Arslan, Hakan, Tasan, Mehmet, Yildirim, Demet, Koksal, Eyüp Selim, Cemek, Bilal
• Format: Journal Article
• Language: English
• Published: Cham Springer-Verlag 01.08.2014; Springer International Publishing; Springer Nature B.V
• Subjects: Aluminum Silicates - chemistry; Atmospheric Protection/Air Quality Control/Air Pollution; Calibration; Clay; correlation; Discriminant analysis; Earth and Environmental Science; Ecology; Ecotoxicology; Environment; Environmental Management; Environmental monitoring; Environmental Monitoring - methods; equations; Field capacity; Irrigation; Laboratories; Mean square errors; Methods; Models, Statistical; Monitoring/Environmental Analysis; Multivariate Analysis; new combination; Physical properties; prediction; Reflectance; Regression analysis; Sand; Silicon Dioxide - chemistry; Silt; Soil - chemistry; Soil analysis; Soil contamination; Soil physical properties; soil sampling; Spectroscopy; Spectrum Analysis; Studies; Wavelengths; Wilting; Wilting point; Soil texture; First order derivatives; Wilting point; Spectral reflectance; Field capacity
• ISSN: 0167-6369; 1573-2959
• DOI: 10.1007/s10661-014-3761-2

In this study, we examined the ability of reflectance spectroscopy to predict some of the most important soil parameters for irrigation such as field capacity (FC), wilting point (WP), clay, sand, and silt content. FC and WP were determined for 305 soil samples. In addition to these soil analyses, clay, silt, and sand contents of 145 soil samples were detected. Raw spectral reflectance (raw) of these soil samples, between 350 and 2,500-nm wavelengths, was measured. In addition, first order derivatives of the reflectance (first) were calculated. Two different statistical approaches were used in detecting soil properties from hyperspectral data. Models were evaluated using the correlation of coefficient (r), coefficient of determination (R ²), root mean square error (RMSE), and residual prediction deviation (RPD). In the first method, two appropriate wavelengths were selected for raw reflectance and first derivative separately for each soil property. Selection of wavelengths was carried out based on the highest positive and negative correlations between soil property and raw reflectance or first order derivatives. By means of detected wavelengths, new combinations for each soil property were calculated using rationing, differencing, normalized differencing, and multiple regression techniques. Of these techniques, multiple regression provided the best correlation (P < 0.01) for selected wavelengths and all soil properties. To estimate FC, WP, clay, sand, and silt, multiple regression equations based on first(2,310)-first(2,360), first(2,310)-first(2,360), first(2,240)-first(1,320), first(2,240)-first(1,330), and raw(2,260)-raw(360) were used. Partial least square regression (PLSR) was performed as the second method. Raw reflectance was a better predictor of WP and FC, whereas first order derivative was a better predictor of clay, sand, and silt content. According to RPD values, statistically excellent predictions were obtained for FC (2.18), and estimations for WP (2.0), clay (1.8), and silt (1.63) were acceptable. However, sand values were poorly predicted (RDP = 0.63). In conclusion, both of the methods examined here offer quick and inexpensive means of predicting soil properties using spectral reflectance data.