Evaluating Machine Learning Algorithms for Onion Mapping in Nueva Ecija, Philippines Using Sentinel-2 Imagery

Abstract. High-value crops like onion are vulnerable to price fluctuations for several reasons, including production shortage, infestation, inflation, importation-related issues, and climate impacts, resulting in high risk for local farmers. Accurate mapping and monitoring can be invaluable in managing these price fluctuations and ensuring long-term stable supply chains, as they enable detailed crop monitoring and yield estimation for onions. In this work, we utilized Sentinel-2 multispectral imagery for onion mapping, applying machine learning algorithms (MLAs) such as Extreme Gradient Boosting Machine (XGBoost), Light Gradient Boosting Machine (LightGBM), and CatBoost Classifier. The input data for the analysis included the 10 RGB, VRE, NIR, and SWIR bands of Sentinel-2 as well as 25 biophysical indices and terrain variables. These indices encompass key indicators for monitoring crop health and suitability like overall vegetation health, chlorophyll content, nitrogen content, soil moisture, soil salinity, soil clay content, Leaf Area Index (LAI), etc. The results showed that among the MLAs tested, CatBoost achieved the highest accuracy (90.0 %), followed by LightGBM (86.7 %), and XGBoost (84.7 %). Among the bands and indices used, the Clay Minerals Ratio (CMR) and Modified Photochemical Reflectance Index (PRI) were consistently identified as the most important features, strongly suggesting that onions are distinguished based on a combination of soil properties and canopy pigment traits.