Assessment of Soil Degradation through Mid-Infrared Spectroscopy and Machine Learning: Predicting Soil Organic Carbon in Fine Fractions of a Desertified Watershed

- 336547
Posters
Favorite this paper
How to cite this paper?
Abstract

This study evaluates Mid-Infrared (MIR) spectroscopy combined with multivariate modeling to estimate total organic carbon (TOC) in a desertified watershed in the Brazilian semiarid region. Understanding carbon dynamics is vital for assessing land degradation and erosion under desertification pressure. Samples from various land-use systems were sieved to <63 µm to represent fractions most susceptible to fluvial transport, matching a monitored mean sediment diameter of 53 µm. Reference TOC analysis revealed high spatial variability, ranging from 4.01 to 56.86 g kg⁻¹ (mean = 25.61 g kg⁻¹; CV = 58.72%). FTIR raw spectra (4000–400 cm⁻¹) were subjected to two treatments: mean-centering and Savitzky–Golay smoothing, aiming to reduce spectral fluctuations and enhance the signal-to-noise ratio. Models were developed using Partial Least Squares Regression (PLSR) and Random Forest (RF), calibrated on 70% of the samples and validated on 30%. Results show that preprocessing significantly influenced performance. While PLSR reached moderate stability (R² = 0.54; RPD = 1.43), it yielded inconsistent negative values in some treatments. RF outperformed PLSR by capturing non-linear spectral responses, with the best result achieved using Savitzky–Golay smoothing combined with RF (R² = 0.81; RMSE = 10.56 g kg⁻¹; RPD = 1.42). These findings highlight MIR spectroscopy coupled with ensemble learning as a robust, non-destructive approach for monitoring soil carbon in sediment-prone fractions, supporting conservation and desertification mitigation strategies in semiarid watersheds.

Share your ideas or questions with the authors!

Did you know that the greatest stimulus in scientific and cultural development is curiosity? Leave your questions or suggestions to the author!

Sign in to interact

Have a question or suggestion? Share your feedback with the authors!

Institutions
  • 1 Universidade Federal Rural de Pernambuco
  • 2 Universidade Federal Rural de Pernambuco | (Federal Rural University of Pernambuco)
Track
  • SOM modeling in agricultural and natural ecosystems
Keywords
Soil organic carbon
Machine Learning
Mid-infrared spectroscopy
Desertification
Fine soil fractions