Machine Learning-Based Alfalfa Height Estimation Using Sentinel-2 Multispectral Imagery.

Bahrami, Hazhir ORCID: https://orcid.org/0000-0003-1371-7337; Chokmani, Karem ORCID: https://orcid.org/0000-0003-0018-0761; Homayouni, Saeid ORCID: https://orcid.org/0000-0002-0214-5356; Adamchuk, Viacheslav I. ORCID: https://orcid.org/0000-0001-7279-3597; Albasha, Rami ORCID: https://orcid.org/0000-0001-5709-3760; Saifuzzaman, Md ORCID: https://orcid.org/0000-0001-8494-8971 et Leduc, Maxime ORCID: https://orcid.org/0000-0001-7097-1636 (2025). Machine Learning-Based Alfalfa Height Estimation Using Sentinel-2 Multispectral Imagery. Remote Sensing , vol. 17 , nº 10. p. 1759. DOI: 10.3390/rs17101759.

Prévisualisation

PDF - Version publiée Télécharger (5MB) | Prévisualisation

Résumé

Climate change is threatening the sustainability of crop yields due to an increasing frequency of extreme weather conditions, requiring timely agricultural monitoring. Remote sensing facilitates consistent and continuous monitoring of field crops. This study aimed to estimate alfalfa crop height through satellite images and machine learning methods within the Google Earth Engine (GEE) Python API. Ground measurements for this study were collected over three years in four Canadian provinces. We utilized Sentinel-2 data to obtain satellite imagery corresponding to the same timeframe and location as the ground measurements. Three machine learning algorithms were employed to estimate plant height from satellite images: random forest (RF), support vector regression (SVR), and extreme gradient boosting (XGB). The efficacy of these algorithms has been assessed and compared. Several widely used vegetation indices, for instance normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), and normalized difference red-edge (NDRE), were selected and assessed in this study. RF feature importance was utilized to determine the ranking of features from most to least significant. Several feature selection strategies were utilized and compared with the situation where all features are used. We demonstrated that RF and XGB surpassed SVR when assessing test data performance. Our findings showed that XGB and RF could predict alfalfa crop height with an R² of 0.79 and a mean absolute error (MAE) of around 4 cm Our findings indicated that SVR exhibited the lowest accuracy among the three algorithms tested, with R² of 0.69 and an MAE of 4.63 cm. The analysis of important features showed that normalized difference red edge (NDRE) and normalized difference water index (NDWI) were the most important variables in determining alfalfa crop height. The results of this study also demonstrated that using RF and feature selection strategies, alfalfa crop height can be estimated with comparably high accuracy. Given that the models were fully trained and developed in Python (v. 3.10), they can be readily implemented in a decision support system and deliver near real-time estimations of alfalfa crop height for farmers throughout Canada.

Type de document:	Article
Mots-clés libres:	alfalfa; crop height; machine learning; Google Earth Engine; remote sensing; Sentinel-2
Centre:	Centre Eau Terre Environnement
Date de dépôt:	18 juill. 2025 15:00
Dernière modification:	18 juill. 2025 15:00
URI:	https://espace.inrs.ca/id/eprint/16535

Gestion Actions (Identification requise)

Modifier la notice