Beyond Spatial Resolution: Comparing Sentinel-2 and PlanetScope imagery for efficient remote mapping

Murilo Henrique Souza de Oliveira, Marlon Fernandes de Souza, Fernando Kenzo Imami Fugihara, RUBENS AUGUSTO CAMARGO LAMPARELLI

TL;DR: Sentinel-2 outperforms PlanetScope when using traditional ML models and pixel-based classification.

Abstract: The choice of remote sensing imagery is critical for balancing accuracy and computational cost in remote sensing applications. This study compares Sentinel-2 (S2) and PlanetScope (PS) satellite imagery for mapping Agricultural Plastic Structures (APS). We investigated the trade-off between PS's higher spatial resolution and the computational demands associated with its larger data volume, using seven machine learning classifiers. Results show that S2 imagery with 10 spectral bands consistently outperformed PS, achieving, on average, 2.5\% higher overall accuracy. Crucially, PS image classification took nearly 10x longer, primarily due to the input file size. These results reveal a key scalability challenge: the small increases in accuracy from using more detailed data sources and complex models can become limiting when deploying large-scale mapping. We conclude that S2 imagery is a suitable source for cost-effective, scalable APS mapping using traditional ML classifiers. Future work should explore deep learning models on these image datasets.

Submission Number: 37