[Paper Review] EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification

Patrick Helber, Benjamin Bischke+2 more|arXiv (Cornell University)|Aug 31, 2017

Remote-Sensing Image Classification62 citations

TL;DR

Introduces the EuroSAT dataset of 27,000 georeferenced Sentinel-2 image patches across 10 classes and benchmarks CNNs achieving up to 98.57% accuracy on RGB data.

ABSTRACT

In this paper, we address the challenge of land use and land cover classification using Sentinel-2 satellite images. The Sentinel-2 satellite images are openly and freely accessible provided in the Earth observation program Copernicus. We present a novel dataset based on Sentinel-2 satellite images covering 13 spectral bands and consisting out of 10 classes with in total 27,000 labeled and geo-referenced images. We provide benchmarks for this novel dataset with its spectral bands using state-of-the-art deep Convolutional Neural Network (CNNs). With the proposed novel dataset, we achieved an overall classification accuracy of 98.57%. The resulting classification system opens a gate towards a number of Earth observation applications. We demonstrate how this classification system can be used for detecting land use and land cover changes and how it can assist in improving geographical maps. The geo-referenced dataset EuroSAT is made publicly available at https://github.com/phelber/eurosat.

Motivation & Objective

• Introduce a large-scale, freely available land use/land cover dataset based on Sentinel-2 imagery.
• Provide multi-spectral benchmarks (13 bands) for deep learning models.
• Evaluate band-specific and band-combined inputs to identify the most informative spectral information.
• Demonstrate baseline applicability for land change detection and map updating.
• Release a georeferenced version to enable broad machine learning research on EO data.

Proposed method

• Assemble 27,000 labeled, geo-referenced 64x64 patches from Sentinel-2A data across 34 European countries.
• Provide RGB and multi-spectral (13-band) versions of EuroSAT for benchmarking.
• Fine-tune and compare CNN architectures (GoogLeNet and ResNet-50) pretrained on ImageNet (ILSVRC-2012) for land use/land cover classification.
• Evaluate single-band versus band-combined inputs, including RGB, CI (color-infrared), and SWIR configurations.
• Analyze per-class confusion and provide a band-importance assessment to determine the most informative spectral bands.

Experimental results

Research questions

• RQ1 What classification accuracy can state-of-the-art CNNs achieve on the EuroSAT dataset with RGB versus multi-spectral inputs?
• RQ2 How do the 13 Sentinel-2 spectral bands perform individually and in combinations for land use/land cover classification?
• RQ3 Is pretraining on ImageNet beneficial for EuroSAT fine-tuning compared to training from scratch?
• RQ4 Can the dataset support land use/land cover change detection and map updating applications?

Key findings

• CNN benchmarks (ResNet-50 and GoogLeNet) achieve state-of-the-art performance on EuroSAT, with ResNet-50 reaching 98.57% accuracy on RGB inputs (80/20 split).
• RGB band combination outperforms single bands and other band combinations, achieving 98.57% (RGB) vs 98.30% (CI) and 97.05% (SWIR) when using ResNet-50.
• Single-band experiments show RGB generally superior; some bands like Red Edge 1 (B05) and SWIR 2 (B12) can yield strong results when upsampled to 10 m.
• CNNs outperform Bag-of-Visual-Words baselines and shallower networks across datasets; pretrained networks provide about 2% higher accuracy on EuroSAT.
• The dataset enables change detection and mapping assistance by applying patch-level classifications to sliding-window analyses of time-separated imagery.