Digital RGB photography and visible-range spectroscopy for soil composition analysis

Publication date: 1 March 2018
Source:Geoderma, Volume 313
Author(s): Matt Aitkenhead, Clare Cameron, Graham Gaskin, Bastien Choisy, Malcolm Coull, Helaina Black
We demonstrate calibration models developed for a number of soil variables, based on colour information derived from standard digital photography and a prototype visible-wavelength spectroscopy device. The samples used were taken from the National Soils Inventory of Scotland and included topsoil samples taken either from the upper horizon or bulked from the top 20cm. A total of 31 variables were investigated, including organic matter content (Loss On Ignition), carbon, nitrogen, pH, sand/silt/clay and a number of elemental concentrations. Model calibrations were developed using visible-wavelength colour information alone and in combination with spatial covariates (topography, climate, soil, vegetation, parent material). A set of thresholds for characterising quality of model calibration was defined using r-squared value, with Good (>0.8), Fair (0.6–0.8), Moderate (0.4–0.6) and Poor (<0.4). For the site descriptors alone, no variables were given Good estimation but several were given Fair, while digital photography alone gave Good estimation for exchangeable H and Fair for Loss on Ignition and C. A combination of site descriptors and colour photography gave Good estimation for four variables and Fair for another six. Visible range spectroscopy alone gave Good estimation for one variable and Fair for another four. Visible-range spectroscopy plus spatial covariates gave the best results, with Good estimation for five variables and Fair for eight. The results indicate that depending on the soil variable of interest, calibration models can be developed for regional soil variable estimation using an appropriate selection of measurement device and data integration. Evaluation of the trained neural network models provided information on the relative importance of individual input variables, allowing us to further identify those inputs of particular interest in this kind of work and to gain a better understanding of the processes linking soils and their environment.