New regularization method advances contrastive self-supervised learning for multispectral remote sensing
BIFOLD’s research group Big Data Analytics for Earth Observation will represent the paper “Rank-based Geographical Regularization: Revisiting Contrastive Self-Supervised Learning for Multispectral Remote Sensing Imagery” at the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2026). The conference takes place from March 6 to 10, 2026, in Tucson, Arizona (USA), an international conference for applied computer vision. WACV focuses on application-driven research at the intersection of computer vision and real-world system requirements.
The paper, authored by Tom Burgert, Leonard Hackel, Paolo Rota, and Begüm Demir, introduces GeoRank, a novel regularization method for contrastive self-supervised learning (SSL) applied to multispectral remote sensing imagery. Regularization guides a model toward desired properties during training by adding constraints to the learning process. Contrastive SSL enables models to learn meaningful representations from unlabeled data by pulling similar samples closer together and pushing dissimilar ones apart in the feature space.
Applied to multispectral satellite imagery, which captures data across multiple wavelength ranges and varies significantly by location, time, and sensor, this approach addresses key challenges in large-scale Earth observation. GeoRank optimizes spherical distances to embed geographical relationships directly into the learned feature space. The method outperforms existing approaches that integrate geographical metadata and consistently improves diverse SSL algorithms such as BYOL and DINO. In addition, the paper provides a systematic investigation of data augmentations, dataset cardinality, and the task dependency of temporal views.
The presentation takes place on Monday, March 9, 2026, between 1:30 and 2:30 pm in Oral Session 5B: Remote Sensing and Sensors.
Paper details:
- Title: Rank-based Geographical Regularization: Revisiting Contrastive Self-Supervised Learning for Multispectral Remote Sensing Imagery
- Authors: Tom Burgert, Leonard Hackel, Paolo Rota, Begum Demir
- Paper: https://arxiv.org/abs/2601.02289