Spectral-Temporal Feature Analysis for Vegetation Classification using Sentinel-2: A Comparison of XGBoost, Random Forest, and Neural Network
Date Issued
October 28, 2025
DOI
10.1117/12.3069699
Abstract
Vegetation classification plays a critical role in land use management, ecological monitoring, and biodiversity conservation. This study investigates the use of Sentinel-2 multispectral time series data and feature importance analysis for grassland type classification, focusing on meadows and orchards. We compare three supervised learning algorithms—Random Forest, XGBoost, and neural network—across four classification tasks: intra-class (standard vs. wet meadows; traditional vs. intensive orchards), binary inter-class (meadows vs. orchards), and multiclass classification. All three models are trained on 156-dimensional pixel-level spectral time series from 2017-2018 and evaluated using 10-fold cross-validation. Our results show that XGBoost achieves the highest performance in binary and multiclass settings, while Random Forest shows high effectiveness for intra-class classification. Additionally, band-wise importance analysis reveals that the NIR band contributes most in intraclass tasks, whereas all bands are more evenly relevant in inter-class tasks. These findings support interpretable, fine-grained vegetation mapping using time series remote sensing data.
File(s)![Thumbnail Image]()
Name
Spectral-Temporal Feature Analysis for Vegetation Classification using Sentinel-2 A Comparison of XGBoost, Random Forest, and Neural N.pdf
Size
443.15 KB
Format
Adobe PDF
Checksum (MD5)
5ee81afcd842f838b11dc94bc6afe267