文章详情

现代压缩方法重新定义了我们处理和分析卫星图像的方式。 In this article, we introduce the 2025 CVPR EARTHVISION Data Challenge — an initiative by the Horizon Europe Embed2Scale ¹ consortium to advance neural compression for Earth Observation data.

有关该主题的全面审查，请阅读我们的最新出版物，用于地理空间分析的有损神经压缩：审查。

Earth Observation (EO) data is growing at an unprecedented pace, with satellites capturing terabytes of high-resolution imagery daily, and EO repositories ranking among the largest data stores globally.例如，近年来，Sentinel数据访问系统在下载中记录了586个PIB。 Efficiently transmitting, storing, and analyzing these vast datasets is essential to unlock insights across diverse applications such as environmental monitoring and disaster management.

The broad spectrum of EO data types — including radar, LiDAR, hyperspectral, and multispectral imagery — presents unique challenges for image compression techniques.当前，应用了经常用于JPEG2000（例如JPEG2000）的手工制作的压缩算法。最近以数据驱动的方法（例如神经压缩）表明有可能达到明显更高的压缩率。

Neural compression typically employs an encoder-decoder architecture trained to convert raw input data into a compact representation, named embedding, from which the original input can be (approximately) reconstructed by a trained decoder.编码器学会生成根据相应数据的特定结构量身定制的紧凑型嵌入。 These embeddings not only reduce the overall data size but also preserve critical features necessary for further analysis, such as scene classification and semantic segmentation.现代下游任务模型无需重建输入，而是可以直接在这些紧凑的压缩表示方面进行培训。

作为Embed2scale财团的一部分，IBM研究正在探索EO的神经数据压缩。 Embed2Scale项目着重于开发紧凑，可重复使用的数据嵌入，并在现实世界应用程序上进行基准测试。

On March 10, 2025, the Embed2Scale consortium launched the 2025 CVPR EARTHVISION Data Challenge , inviting researchers and AI practitioners to develop innovative EO data compression techniques. Participants are tasked with designing an encoder that transforms high-dimensional satellite image data cubes into compact, fixed-size embeddings.

During the development phase in March, participants will pretrain their encoders using self-supervised learning methods that underpin neural compression and EO foundation models. The SSL4EO-S12 ⁴ dataset combines radar and multispectral image bands from four-season snapshots, illustrating the multi-temporal and multi-modal character of EO data.这些数据包必须被约7000倍压缩到紧凑的1024维嵌入中。

For evaluation, participants will submit embeddings generated with their compression encoder, which will be assessed through hidden downstream tasks to determine how effectively relevant information is preserved.

The challenge provides participants with the opportunity to test and enhance state-of-the-art compression methods, and highlights the benefits of compact embeddings. Beyond advancing lossy neural compression, the challenge promotes open-source and open-science efforts and establishes a benchmarking framework for compressed geospatial embeddings. The Embed2Scale consortium invites the AI community to adopt this novel data benchmark evaluation scheme and contribute additional downstream tasks.挑战的结果，代码和数据将在今年的CVPR会议上介绍。

挑战是开放的，可以通过我们的eval.ai挑战门户访问。