We looked at the MultispeQ's ability to predict total carbon content and black carbon content at sites in Kenya with 99 total soil samples, and used random forest and ensemble partial least square regressions for models predicting carbon content from MultispeQ responses. The predictions were made for total carbon and black carbon using both MultispeQ response alone and including sample descriptions like site, depth, and year with the MultispeQ response. All carbon types had better predictions including other factors like site and year in addition to wavelengths, which may indicate value in site-specific calibrations.
The carbon mixtures of activated carbon (analogous to black carbon), cellulose, and lignin in silica were able to give a better picture of what the MultispeQ actually detects. Activated carbon was able to be predicted using the MultispeQ responses and indicates the device can directly measure actived carbon/black carbon. Cellulose and lignin carbon sources did not have high correlations to the reflectance responses. These carbon sources were also difficult to create a general model and had issues with overfitting.
Further testing and more diverse compositions may improve cellulose, lignin, and total carbon correlation to wavelength responses.
Random forest used with actived carbon produced consistent results when comparing the model performance to training and test samples, and had the added benefit of being able to rank wavelengths.
Important Wavelengths Total Carbon (Soil/ Artificial Mix)
650 nm / 650nm
605 nm / 940nm
530nm / 605nm
Black Carbon (Soil/ Artificial Mix)
720nm / 850nm
940nm / 650nm
530nm / 940nm