BigEarthNet: A Large-Scale Benchmark Archive For Remote Sensing Image Understanding (1902.06148v3)

Abstract: This paper presents the BigEarthNet that is a new large-scale multi-label Sentinel-2 benchmark archive. The BigEarthNet consists of 590,326 Sentinel-2 image patches, each of which is a section of i) 120x120 pixels for 10m bands; ii) 60x60 pixels for 20m bands; and iii) 20x20 pixels for 60m bands. Unlike most of the existing archives, each image patch is annotated by multiple land-cover classes (i.e., multi-labels) that are provided from the CORINE Land Cover database of the year 2018 (CLC 2018). The BigEarthNet is significantly larger than the existing archives in remote sensing (RS) and thus is much more convenient to be used as a training source in the context of deep learning. This paper first addresses the limitations of the existing archives and then describes the properties of the BigEarthNet. Experimental results obtained in the framework of RS image scene classification problems show that a shallow Convolutional Neural Network (CNN) architecture trained on the BigEarthNet provides much higher accuracy compared to a state-of-the-art CNN model pre-trained on the ImageNet (which is a very popular large-scale benchmark archive in computer vision). The BigEarthNet opens up promising directions to advance operational RS applications and research in massive Sentinel-2 image archives.

• The paper introduces BigEarthNet, a comprehensive benchmark featuring 590,326 Sentinel-2 image patches with multi-label annotations from the CORINE Land Cover database.
• It demonstrates significant improvements in classification accuracy using a custom S-CNN model trained on the full spectral dataset compared to RGB-only and ImageNet pre-trained models.
• The dataset’s extensive spectral information and detailed annotations pave the way for advanced remote sensing applications such as land cover mapping and environmental monitoring.

BigEarthNet: Enhancing Remote Sensing Through Large-Scale Image Data

The paper "BigEarthNet: A Large-Scale Benchmark Archive for Remote Sensing Image Understanding" introduces a significant resource in the domain of remote sensing (RS) and machine learning. BigEarthNet is a comprehensive dataset composed of 590,326 Sentinel-2 image patches, aimed at addressing several critical limitations faced by the RS community in leveraging deep learning techniques. The dataset is noteworthy for its scale and multi-label annotations, which are derived from the CORINE Land Cover database (CLC 2018). This allows a more nuanced understanding of RS images compared to existing datasets that typically provide only single-label annotations.

Overview of Contributions

The BigEarthNet archive is distinctive for several reasons:

1. Scale and Coverage: Comprising over half a million image patches, BigEarthNet is far larger than prior RS benchmarks, making it suitable for training deep learning models that require extensive data.
2. Multi-Label Annotations: Each image patch in BigEarthNet is annotated with multiple land-cover classes, accommodating the inherent complexity and diversity in RS imagery. This contrasts with the predominantly single-label annotations found in other datasets, providing a richer training dataset for developing models capable of semantic segmentation.
3. Spectral Information: The dataset includes bands from Sentinel-2’s 13 spectral bands (excluding the 10\textsuperscript{th} band due to lack of surface information), offering comprehensive spectral data that enhances the potential for detailed analysis over the traditional RGB channels.

Experimental Analysis

The authors conducted an empirical analysis demonstrating the utility of BigEarthNet by comparing a shallow Convolutional Neural Network (CNN) trained from scratch on the dataset against the fine-tuned Inception-v2 model pre-trained on the ImageNet dataset. Results showed significant improvements in classification accuracy when using models trained on BigEarthNet. Specifically, the S-CNN model, utilizing all spectral bands, achieved a substantial performance advantage over both the S-CNN using only RGB channels and the fine-tuned Inception-v2 model. This highlights the inadequacies of using models trained on general computer vision datasets for specialized RS tasks.

Implications for Remote Sensing

The introduction of BigEarthNet represents a substantial step towards overcoming data resource limitations in RS. It provides a robust benchmark for the development of more sophisticated models tailored to the specific characteristics of RS data. The multi-label nature of the dataset encourages the development of models that can handle more complex classification tasks, reflecting the real-world scenarios where multiple land-cover classes coexist.

Future Directions

The research community can leverage BigEarthNet to train and evaluate algorithms aimed at various RS applications, such as land cover mapping and environmental monitoring. Additionally, the authors suggest that extending BigEarthNet by incorporating more data will further advance the field. The potential for employing more advanced neural architectures, potentially with residual connections and custom designs for RS data, holds promise for pushing forward the state-of-the-art in RS image analysis.

In conclusion, BigEarthNet is positioned to become an invaluable asset for the remote sensing community, presenting rich opportunities for advancements in both theoretical research and practical applications. Its large scale and detailed annotations challenge researchers to shift from conventional single-label, RGB model constraints and explore the full spectrum of possibilities presented by RS imaging.