# Prediction of vegetation indices from down-sampled hyperspectral data using machine learning: A novel framework for olive crop monitoring

## TL;DR

This paper introduces a machine learning framework to predict plant health indicators from low-resolution spectral data, enabling cost-effective crop monitoring.

## Contribution

A novel machine learning framework for predicting vegetation indices from down-sampled hyperspectral data is proposed.

## Key findings

- High prediction accuracy (R² up to 0.99) was achieved for key vegetation indices using 100 nm spectral resolution.
- The framework enables accurate plant health monitoring with significantly reduced spectral data.
- The method is particularly effective for crops like olives, where water and nutrient status are critical.

## Abstract

Accurate plant health monitoring relies on hyperspectral imagery to extract vegetation spectral signatures and compute vegetation indices (VIs), which are critical for phenotyping and crop condition assessment. However, the requirement for high spectral resolution significantly increases the cost and complexity of data acquisition. In this study, we proposed a novel machine learning-based framework for predicting VIs from down-sampled hyperspectral reflectance data. The aim was to reduce the dependency on high-resolution spectral imagery without compromising prediction accuracy. The framework integrated correlation-based feature selection with four regression models to identify and utilize the most informative spectral bands from coarsely sampled data. The system was trained and validated using a data set consisting of 555 spectral signatures collected from olive leaves at five stages of dehydration, with spectral resolutions ranging from 1 to 100 nm. A total of 25 vegetation indices, commonly used in the estimation of water stress, chlorophyll, and nitrogen, were predicted on various sampling scales. Experimental results show that even with 100 nm spectral resolution, the proposed framework achieves high prediction accuracy, with coefficients of determination reaching 0.99 for RVSI, VOPT, and SPADI indices. These findings demonstrate that accurate vegetation index estimation is achievable with significantly fewer spectral bands, offering a cost-effective solution for large-scale plant health monitoring. This framework lays the groundwork for the development of low-cost, data-efficient remote sensing systems for precision agriculture, especially in crops such as olives, where health dynamics are sensitive to water and nutrient status.