Predicting soil organic carbon in a small farm system using in situ spectral measurements and the random forest regression
A research report submitted to the Faculty of Science, University of the Witwatersrand, Johannesburg, in partial fulfillment of the requirements for the degree of Master of Science (Geographical Information Sciences and Remote Sensing) Johannesburg, 2017 === Soil organic carbon is considered as the...
| Main Author: | |
|---|---|
| Format: | Others |
| Language: | en |
| Published: |
2017
|
| Subjects: | |
| Online Access: | Bangelesa, Freddy Fefe (2017) Predicting soil organic carbon in a small farm system using in situ spectral measurements and the random forest regression, , University of the Witwatersrand, Johannesburg, <http://hdl.handle.net/10539/23485> http://hdl.handle.net/10539/23485 |
| Summary: | A research report submitted to the Faculty of Science, University of the Witwatersrand, Johannesburg, in partial fulfillment of the requirements for the degree of Master of Science (Geographical Information Sciences and Remote Sensing)
Johannesburg, 2017 === Soil organic carbon is considered as the most determining indicator of soil fertility. The purpose of this research was to predict the soil organic carbon in the Mokhotlong region, eastern of Lesotho using in situ spectral measurements and random forest regression. Soil reflectance spectra were acquired by a portable field spectrometer.
The performance of random forest regression was assessed by comparing it with one of the most popular models in spectroscopy, partial least square regression. Laboratory spectroscopy measurements of the soil samples were analysed for assessing the accuracy of in situ spectroscopy based-models. The effect of the Savitzky−Golay first derivative in improving partial least square regression and random forest regression in both spectral data was also assessed.
The results indicated that the random forest regression could accurately predict the soil organic carbon contents on an independent dataset using in situ spectroscopy data (RPD = 3.77, Rp2= 0.88, RMSEP = 0.64%). The overall best predictive model was achieved with the derivative laboratory spectral data using random forest with the optimum number of key wavelengths (RPD = 3.77, Rp2= 0.88, RMSEP = 0.64%). In contrast, partial least square regression was likely to overfit the calibration dataset. Important wavelengths to predict soil organic contents were localised around the visible range (400-700 nm). An implication of this research is that soil organic carbon can accurately be estimated using derivative in situ spectroscopy measurements and random forest regression with key wavelengths. === MT 2017 |
|---|