# Predicting with limited data - Increasing the accuracy in VIS-NIR diffuse reflectance spectroscopy by SMOTE

## TL;DR

This paper introduces a framework using SMOTE to improve soil property predictions from diffuse reflectance spectroscopy when field spectra are scarce, significantly enhancing model accuracy.

## Contribution

The study presents a novel application of SMOTE to calibrate PLS models with limited field spectra, improving prediction accuracy in VIS-NIR spectroscopy.

## Key findings

- Root mean-squared error decreased from 6.18 to 2.12 mg g$^{-1}$.
- $R^2$ increased from -0.53 to 0.82.
- Model performance was substantially improved with synthetic field spectra.

## Abstract

Diffuse reflectance spectroscopy is a powerful technique to predict soil properties. It can be used in situ to provide data inexpensively and rapidly compared to the standard laboratory measurements. Because most spectral data bases contain air-dried samples scanned in the laboratory, field spectra acquired in situ are either absent or rare in calibration data sets. However, when models are calibrated on air-dried spectra, prediction using field spectra are often inaccurate. We propose a framework to calibrate partial least squares models when field spectra are rare using synthetic minority oversampling technique (SMOTE). We calibrated a model to predict soil organic carbon content using air-dried spectra spiked with synthetic field spectra. The root mean-squared error of prediction decreased from 6.18 to 2.12 mg g$^{-1}$ and $R^2$ increased from $-$0.53 to 0.82 compared to the model calibrated on air-dried spectra only.