SSL4Eco: A Global Seasonal Dataset for Geospatial Foundation Models in Ecology
Paper | Webpage | Code
Elena Plekhanova, Damien Robert, Johannes Dollinger, Emilia Arens, Philipp Brun, Jan Dirk Wegner, Niklaus Zimmermann,
CVPR EarthVision workshop 2025 Poster 🖼

This work proposes a spatio-temporal sampling recipe for building pretraining sets for geospatial foundation models targeted for downstream macroecological applications. Simply put, we sample geolocations uniformly across the landmass and sample dates following local, phenology-informed seasons. We create a dataset SSL4Eco following this sampling recipe, on which we pretrain a model following the existing Seasonal Contrast approach. We show that our simple dataset construction recipe significantly impacts the performance on diverse downstream macroecological tasks, compared to other datasets, and pretrained models.

Climplicit: Climatic Implicit Embeddings for Global Ecological Tasks
Paper | Webpage | Code
Johannes Dollinger, Damien Robert, Elena Plekhanova, Lukas Drees, Jan Dirk Wegner,
ICLR Tackling Climate Change with Machine Learning workshop 2025 Poster 🖼

Climplicit proposes to learn global, multi-temporal, implicit climatic representations to facilitate the use of climatic data 🌦 for downstream applications with minimal compute, memory, and deep learning expertise requirements 🦋.

GSR4B: Biomass Map Super-Resolution with Sentinel-1/2 Guidance
Paper | Webpage | Code
Kaan Karaman, Yuchang Jiang, Damien Robert, Vivien Sainte Fare Garnot, Maria João Santos, Jan Dirk Wegner
ISPRS Geospatial Week 2025 Oral 🎤

GSR4B proposes a method for producing high-resolution above-ground biomass 🪵🌳 maps using guided super-resolution. More specifically, we show that it is possible to upsample 100m-resolution AGB maps to 10m resolution with the help of Sentinel-1/2 imagery 🛰. The resulting estimates are more accurate than direct regression solely from Sentinel imagery and less prone to over-estimating low values and under-estimation high values.

Lossy Neural Compression for Geospatial Analytics: A Review
Paper
Carlos Gomes, Isabelle Wittmann, Damien Robert, Johannes Jakubik, Tim Reichelt, Michele Martone, Stefano Maurogiovanni, Rikard Vinge, Jonas Hurst, Erik Scheurer, Rocco Sedona, Thomas Brunschwiler, Stefan Kesselheim, Matej Batic, Philip Stier, Jan Dirk Wegner, Gabriele Cavallaro, Edzer Pebesma, Michael Marszalek, Miguel A Belenguer-Plomer, Kennedy Adriko, Paolo Fraccaro, Romeo Kienzler, Rania Briq, Sabrina Benassou, Michele Lazzarini, Conrad M Albrecht
IEEE Geoscience and Remote Sensing Magazine 2025

This paper reviews the state of lossy neural compression 🤖🗜️, with a focus on its application and potential for the manipulation and analysis of Earth Observation data 🌍 such as satellite imagery 🛰 and climate model predictions 🌦.

Scalable 3D Panoptic Segmentation as Superpoint Graph Clustering
Damien Robert, Hugo Raguet, Loïc Landrieu
Paper | Webpage | Code
3DV 2024 Oral 🎤 (top 5.3% submissions)

SuperCluster proposes a framework for efficient panoptic segmentation of large-scale point clouds. We formulate the panoptic task as a graph clustering problem. We train a model to predict desirable node and edge attributes to be used as input for a downstream graph clustering algorithm. This allows for training a model with only local supervision, without the need for non-maximum suppression, instance matching, and without any prerequisite on the number of objects in the scene. SuperCluster achieves new SOTA panoptic segmentation on indoor datasets S3DIS Area 5 (50.1 PQ (+7.8)) and ScanNetV2 (58.7 PQ (+25.2)), as well as outdoor datasets KITTI-360 (48.3 PQ) and DALES (61.2 PQ).
🦋 210k param. | ⚡ Train S3DIS F5 in 4h | 💾 20M-point inference on 1 GPU

Efficient 3D Semantic Segmentation with Superpoint Transformer
Damien Robert, Hugo Raguet, Loïc Landrieu
Paper | Webpage | Code
ICCV 2023 Poster 🖼 (top 26.8% submissions)

SPT is a superpoint-based transformer 🤖 architecture that efficiently ⚡ performs semantic segmentation on large-scale 3D scenes. This method includes a fast algorithm that partitions 🧩 point clouds into a hierarchical superpoint structure, as well as a self-attention mechanism to exploit the relationships between superpoints at multiple scales. We reach SOTA on S3DIS 6-Fold (76.0 mIoU), KITTI-360 Val (63.5 mIoU), and DALES (79.6 mIoU)n with:
🦋 212k param. | ⚡ Train S3DIS F5 in 3h | ⌚ SPG preprocessing ÷7

Learning Multi-View Aggregation In the Wild for Large-Scale 3D Semantic Segmentation
Damien Robert, Bruno Vallet, Loïc Landrieu
Paper | Webpage | Code | Video
CVPR 2022 Oral 🎤 and Best Paper finalist 🎉 (top 0.4% submissions)

An end-to-end multi-view aggregation method for 3D semantic segmentation from images and point clouds. We reach SOTA on S3DIS and KITTI-360 without requiring point cloud colorization, meshing, or depth sensors: just point clouds ☁, images 📸, and their poses.