The workshop on global datasets and models for Earth observation returns at ECCV 2026 — this September in Malmö, Sweden.
Earth observation (EO) presents unique challenges and opportunities that set it apart from other fields of machine learning and computer vision. EO data is abundant, repeatedly covering a large but bounded environment — our planet — and the co-location and evolution of these observations is a rich, multimodal, multitemporal source of information. Yet the distribution of EO data is non-stationary and spatially biased, with large parts of the world strongly under-represented.
TerraBytes is an initiative to address these challenges at the intersection of data curation, data archiving, and representation learning, fostering a holistic discussion that covers every step from downlinked satellite data to training paradigms and downstream applications.
After a successful first edition at ICML 2025, TerraBytes returns bigger at ECCV 2026. The programme and the 36 accepted papers are published below; remaining details will be confirmed on this page as they are settled.
Full-day workshop, 8 September 2026. All times are local to Malmö (CEST).
This schedule is preliminary and may still change — including at short notice — due to ECCV 2026 conference organisation. Please check the exact timings on the day of the event via the official ECCV programme and on-site signage.
Asterisk Labs is sponsoring the TerraBytes II Travel Grant Award — helping authors who would otherwise be unable to travel present their work in Malmö.
36 papers were accepted to TerraBytes II, grouped below by presentation format. Select a title to read its abstract. Accepted full-length papers may opt in to the archival proceedings; short papers are non-archival.
Given spotlight talks during the morning and afternoon spotlight sessions, in addition to the poster session.
Satellite imagery differs from natural images in viewpoint, resolution, scale variation, and the prevalence of small objects — demanding both region-level spatial reasoning and holistic scene understanding. Existing remote-sensing approaches are fragmented: dual-encoder retrieval models scale well but cannot interleave modalities, whereas generative assistants support grounding, yet are inefficient for retrieval. Benchmarks mirror this split: interleaved evaluations mainly target generative assistants, while cross-modal retrieval benchmarks target dual encoders. To bridge this gap, we introduce RSMEB, a unified remote sensing benchmark that evaluates cross-modal and interleaved retrieval across 21 tasks under a single ranking protocol, enabling comprehensive comparison of retrieval models on region- and geo-aware capabilities as well as conventional retrieval. As a strong reference baseline, we present VLM2GeoVec, an instruction-conditioned, single-encoder interleaving formulation tailored to remote sensing that packs image, text, bounding-box, and geo-coordinate tokens into one sequence and learns a unified embedding via contrastive training. Across RSMEB, VLM2GeoVec achieves 26.6% P@1 in region-caption retrieval (+25 percentage points), 32.5% in referring-expression retrieval (+19), and 17.8% in semantic geo-aware retrieval (>3× prior best), while remaining competitive in conventional scene classification and text–image retrieval in zero-shot settings. Together, the proposed suite and reference baseline standardize evaluation and deliver a unified embedder for scalable retrieval and region-/geo-aware grounding. Code, model checkpoints, and data will be released upon acceptance.
Operational crop type mapping requires models that generalise across years, resolve fine-grained crop taxonomies, and distinguish cropland from surrounding landscapes. However, existing crop mapping datasets enable evaluation of these requirements only in isolation. We therefore introduce SwissCrop25, a national-scale crop mapping dataset spanning seven growing seasons (2019-2025), combining Sentinel-2 time series, daily temperature observations, a fine-grained 73 crop taxonomy with grassland management types, and 5 explicit non-crop land cover classes. To evaluate realistic deployment conditions, we define a five-fold leave-one-year-out protocol with joint cropland delineation and crop classification for benchmarking representative crop mapping architectures. Evaluating U-TAE, TSViT, and Galileo reveals differences between architectures hidden by conventional benchmarks. In this setting, domain-specific models outperform Galileo, a pretrained Earth observation foundation model, with TSViT achieving the best overall performance and a 12 pp mIoU advantage over U-TAE. TSViT's advantage increases with taxonomic specificity and class rarity, highlighting the importance of fine-grained evaluation. SwissCrop25 also exposes substantial interannual distribution shifts and shows that temperature priors improve robustness. Finally, in-season evaluation reveals a trade-off between models, with U-TAE providing stronger early-season performance on common crops and TSViT gaining an advantage later in the season through improved rare-class discrimination. SwissCrop25 provides a challenging testbed for evaluating crop mapping systems under realistic operational conditions and is publicly released.
Remote sensing single-image super-resolution aims to generate high-resolution imagery from low-resolution observations while preserving fine spatial structures such as roads, building boundaries, field edges, and land-cover transitions. Recent Swin Transformer-based models, including Swin2SR, provide strong spatial context modeling through shifted-window self-attention; however, their feed-forward networks remain generic channel-mixing modules and do not explicitly distinguish between low-frequency structural content and residual detail features. To address this limitation, we propose SFG-SwinSR, a Spatial-Frequency Gated Swin Transformer for cross-sensor remote sensing super-resolution. SFG-SwinSR replaces the standard feed-forward network in each Swin2SR transformer block with a lightweight Spatial-Frequency Gated Feed-Forward Network. The proposed module estimates low-frequency structural content through a depthwise low-pass branch, derives residual detail features by subtraction, refines them using a lightweight spatial branch, and adaptively reinjects useful details through a bottleneck gate. Experiments on three real cross-sensor benchmarks, SEN2VENµS, OLI2MSI, and SEN2NAIP, together with an auxiliary synthetic SpaceNet Challenge 3 setting, show consistent improvements across most evaluation settings and competitive performance against recent Swin-based baselines. The results demonstrate that spatial-frequency feature transformation inside transformer feed-forward networks provides an effective and lightweight inductive bias for structure-aware cross-sensor remote sensing super-resolution.
Accurate refinement of Rational Polynomial Camera (RPC) models is essential for high-quality satellite image geolocation. In ground control point (GCP)-free multi-view pipelines, this refinement is commonly performed through bundle adjustment from automatically extracted image correspondences. However, conventional RPC bundle adjustment pipelines rely on handcrafted feature matching, which becomes unreliable in multi-date collections affected by seasonal, illumination, and land-cover changes. We develop an appearance-aware RPC refinement pipeline by systematically evaluating and integrating learned local matching and global image descriptors, together with a novel similarity-based image-pair selection strategy. This reduces redundant and error-prone matching while preserving the connectivity of the matching graph. Experiments on seasonally diverse WorldView-3 images show that the resulting pipeline improves GCP-free relative RPC refinement over open-source baselines, achieving lower geometric consistency errors while reducing matching time on collections with 39-42 views. These results demonstrate the effectiveness of the proposed system on WorldView-3 imagery and motivate broader evaluation across sensors and datasets.
Remote sensing scene classification is increasingly relying on domain-specific foundation models. Moreover, transductive inference, which exploits the collective statistical structure of the entire unlabeled query set, appears to naturally match remote sensing pipelines where large images are routinely split into patches and inferred as a batch. In this work, we introduce LC-TIM (Locally Consistent Transductive Information Maximization), which extends the state-of-the-art Transductive Information Maximization for Few-Shot CLIP (TIM++) objective with a local consistency regularizer that enforces prediction agreement between each query sample and its κ nearest feature-space neighbors. The regularizer enters as a single multiplicative factor in the closed-form q-update, adding negligible computational overhead. We further propose a multi-source extension that fuses the affinity graph from multiple remote sensing foundation model, further boosting classification accuracy. To assess these methods, we establish the first comprehensive, open-source benchmark for transductive few-shot RS scene classification, evaluating LP++, TransCLIP, TIM++, and LC-TIM across ten diverse datasets, two remote sensing vision-language models, and across various few-shot shot settings. Our experiments show that transductive methods consistently outperform inductive baselines, and that LC-TIM achieves state-of-the-art accuracy, with the largest gains in the low-shot regime where neighborhood cues are most informative. Code will be made available upon acceptance.
We introduce TerraDiT, a diffusion transformer designed for text-to-satellite image generation with point-based control. Existing controlled satellite image generative models often require pixel-level maps that are time-consuming to acquire, yet semantically limited. To address this limitation, we introduce a novel point-based conditioning framework that controls the generation process through the spatial location of the points and the textual description associated with each point, providing semantically rich control signals. This approach enables flexible, annotation-friendly, and computationally simple inference for satellite image generation. To this end, we introduce an adaptive local attention mechanism that effectively regularizes the attention scores based on the input point queries. We systematically evaluate various domain-specific design choices for training TerraDiT, including the selection of satellite image representation for alignment and geolocation representation for conditioning. Our experiments demonstrate that TerraDiT achieves impressive generation performance, surpassing the state-of-the-art remote sensing generative models. Our models, dataset, and code will be made publicly available.
While online Earth monitoring with Satellite Image Time Series (SITS) is essential for tracking rapid anthropogenic changes, existing remote sensing foundation models remain unsuited for streaming SITS. We introduce IRMA, the first foundation model tailored for online land monitoring. IRMA leverages a novel dual-form framework that unifies parallelized multi-modal pre-training (Sentinel-1/2) with an efficient recurrent inference mechanism. Our self-supervised objective produces latent representations that simultaneously maintain temporal stability against seasonal variations and sensitivity to permanent land modifications. To evaluate our method, we present MELBA, a multi-temporal benchmark spanning land-cover, building density, and gold-panning tasks. Performance evaluations demonstrate that IRMA delivers competitive results compared to state-of-the-art baselines while requiring fewer parameters.
Geospatial foundation models aim to learn representations that transfer across regions and sensors, yet evaluating them on specific tasks requires large, high-quality, multi-modal benchmarks that measure how well such models extract value from data. Concerning flood mapping, existing datasets rarely combine bi-temporal SAR and co-registered optical imagery at scale, leaving the value of foundation models for this downstream task largely untested. We introduce GEOID-Flood, a large-scale multi-modal flood segmentation benchmark, derived from Copernicus Emergency Management Service activations, spanning 219 events across 65 countries over ten years. The dataset provides more than 14 000 tiles with co-registered pre- and post-event Sentinel-1, in GRD and RTC format, pre-event Sentinel-2 composite, and DEM, including manually validated labels that separate background from permanent water and flooded water. Using this benchmark, we evaluate foundation models against conventional encoders across single-image, multi-temporal, and multi-modal protocols. We report three main findings: foundation models offer a consistent but modest advantage; optical–SAR fusion with finetuning best resolves transient flooding; and models trained on GEOID-Flood transfer to unseen events better than those trained on existing datasets.
Vision–language models (VLMs) show promise for disaster-response remote sensing, but existing benchmarks mainly emphasize scene-level or damage-centric assessment. To study this building-centric gap, we introduce DisasterInsight, a diagnostic benchmark built on xBD, a pre/post-disaster satellite dataset with building-level damage labels. DisasterInsight enriches building instances with OpenStreetMap-derived functional labels and contains 134,108 task-specific instruction records across 15 task types, spanning instance-level assessment, scene-level counting, multi-instance reasoning, and structured report generation. The benchmark supports RGB pre/post-disaster imagery, single- and multi-view instance formulations, and scene-level RGB/SAR diagnostic inputs. Experiments with general-domain and remote-sensing VLMs show that models perform better on visible damage cues than on building-function understanding, multi-instance reasoning, counting, and grounded reporting. Instruction tuning improves performance on several tasks but does not close this building-centric gap.
Airborne laser scanning (ALS) point clouds are essential for large-scale 3D scene understanding, yet acquiring dense and complete scans remains expensive and operationally demanding. While point cloud upsampling offers a practical alternative, progress on real-world ALS scene upsampling has been hindered by two key challenges: the lack of standardized benchmarks and the difficulty of preserving structural fidelity under sparse observations. To address these limitations, we introduce RAPU, a dedicated benchmark for scene-level ALS point cloud upsampling that provides realistic sparse–dense training pairs constructed from overlapping flight-line acquisitions. We further propose the Geometry-Appearance Conditioned Diffusion Model (GCDM), a conditional diffusion framework that augments sparse-point conditioning with complementary geometric and appearance priors. GCDM combines online PCA-kNN surface-normal features with multi-view projected-image features rendered from ten virtual viewpoints, and integrates them through a lightweight transformer-based cross-modal attention block to guide diffusion-based reconstruction. Experiments on RAPU under x4 upsampling show that GCDM consistently outperforms existing diffusion-based baselines, yielding better geometric fidelity, stronger structural consistency, and more complete scene reconstruction. Code and data will be released upon acceptance.
Multimodal remote sensing image registration is essential for multi-sensor data fusion but remains challenging due to significant radiometric differences across sensors. Most existing deep learning approaches require large-scale, perfectly aligned multimodal datasets, which are costly and difficult to acquire. We introduce ALIGNet, a zero-shot self-supervised framework that achieves state-of-the-art performance on multimodal remote-sensing registration without requiring cross-modal training pairs. Our framework uses a frozen encoder pre-trained on large-scale optical satellite imagery, and trains the alignment module on natural images augmented with sensor-like transformations. This design enables effective generalization to unseen real-world sensors. Experimental results on MDAS and DREAM benchmarks show that ALIGNet consistently outperforms state-of-the-art methods, reducing overall RMSE by 35% on MDAS and 36% on DREAM compared to the second-best models. Notably, our framework demonstrates superior resilience in highly heterogeneous scenarios, such as radar-optical and high- vs low-resolution alignment, while exhibiting the highest bidirectional robustness. ALIGNet provides a practical, scalable solution for real-world remote sensing workflows where co-registered data is scarce.
Accurate 3D reconstruction from satellite imagery typically relies on near-simultaneous stereo pairs, limiting its applicability to diachronic settings where multi-date images exhibit varying seasonal and illumination conditions. Training dense stereo matching models robust to appearance changes is a long-standing challenge, as aligned multi-date imagery and ground-truth geometry are costly to obtain at scale. We propose SeasonStereo, a scalable framework that addresses disparity estimation from diachronic satellite images by training on synthetic image pairs with controlled seasonal appearance variation, while leveraging zero-shot geometric priors from foundation models. SeasonStereo matches the disparity accuracy of state-of-the-art LiDAR-supervised models, while producing sharper geometric details without requiring real aligned multi-date training products or LiDAR-derived labels. As a result, SeasonStereo offers a practical path toward large-scale 3D reconstruction from heterogeneous satellite images with reduced supervision cost.
Open-vocabulary segmentation labels arbitrary categories from a text query without per-class training, yet on remote sensing imagery it underperforms on categories it handles reliably elsewhere. We find that much of this gap traces to the text query rather than to the segmentation model. Because these models are not specialized for overhead imagery, the class name that serves as the query is often a weak address into the vision-language embedding space. We show that a better name repairs part of the gap, while the remaining failures call for an address that no name provides. We recover that address from a few examples through textual inversion on a frozen model, keeping inference text only. On a representative benchmark this raises the mean intersection over union on the affected categories from 3.9 to 39.4, and across eight remote sensing datasets it improves over few-shot methods that instead inject visual prompts at inference.
We present Helios, a family of pretrained remote sensing foundation models built on top of OlmoEarth. This model family is designed to improve OlmoEarth's efficiency, whilst maintaining performance. Compared to OlmoEarth, Helios cuts compute costs during training (1.7× reduction in GPU hours required to train our Base models) and inference (2.9× reductions in MACs on Sentinel-2 tasks), while maintaining OlmoEarth's overall performance. All training code will be shared upon publication.
Self-supervised pretraining on remote sensing imagery typically treats all samples as equally informative, despite large variability in geographic and visual structure. We propose a curriculum learning strategy for self-supervised Earth observation that ranks samples by geographic isolation, a label-free proxy derived entirely from geolocation metadata already present in geospatial datasets, requiring no image decoding, no model feedback, and no manual annotation. Unlike visual complexity proxies, it scales as O(D log D) with dataset size D and is well-defined for both contrastive and reconstructive objectives. We integrate the proposed measure into MoCoV2 and MAE pretraining and evaluate across three downstream tasks from CopernicusBench (BigEarthNet, DFC-2020, LCZ). Our curriculum reaches baseline final-epoch performance using as few as 20% of the training budget (MAE) and at most 40% (MoCo) of the training budget, and improves final downstream performance by up to +5 mAP on BigEarthNet, with gains of 1–5 points across benchmarks, matching visual-complexity curricula while reducing pre-computation cost by more than 140× (4 s vs. 568 s on SSL4EO). A CKA and effective-rank analysis further reveals that curriculum-trained encoders develop higher-dimensional, more uniformly utilized embedding spaces throughout training.
Small satellites are increasingly used for Earth observation, but their limited on-board resources and less precise attitude control make accurate georeferencing a difficult task whose evaluation is not yet standardized at the small-satellite scale. We introduce PhindMe, a difficulty-stratified reference dataset of 144 PhiSat-2 scenes covering eight land-cover types and four difficulty classes, together with an open evaluation framework that benchmarks any external georeferencing method through a unified interface. The reference geometry is obtained by calibrating a rigorous pushbroom sensor model against Sentinel-2 Level-1C tie points extracted with learned matchers (EfficientLoFTR and SuperPoint+LightGlue), fitting Rational Polynomial Coefficient (RPC) models, and validating against Sentinel-2 GRI control points. The framework consumes external RPC outputs, applies a fixed orthorectification and gradient-NCC verification chain, and reports geolocation error in meters and pixels alongside CPU, RAM, and wall-time consumption on both workstation and Jetson Orin Nano hardware. On the 136-scene anomaly-filtered split, our reference pipeline reaches a median RMSE of 7.35 m (1.55 px) and CE90 of 10.04 m (2.11 px), and the Jetson edge baseline runs in 137 s and 1.2 GB peak RSS per scene, demonstrating that the evaluation pipeline is operable within a 15 W edge envelope.
Semantic segmentation of Earth Observation data remains a challenging task. Despite continuous advancements in remote sensing architectures, standard fully supervised pipelines frequently rely on standard decoders and pixel-wise cross-entropy objectives, regardless of often exhibiting intra-image long-range spatial correlations, and long-tailed distribution of classes dominated by background. In this work, we revisit the foundational mechanisms of dense predictions in remote sensing. First, we introduce a novel Graph Attention module embedded within the skip connections, which explicitly aggregates long-range semantic correlations while preserving local spatial proximity. Second, we extend the principles of Balanced Contrastive Learning (BCL), a methodology to address long-tail distributions in image classification, to the dense prediction domain, proposing Pixel-Level Balanced Contrastive Learning (PL-BCL) as an extension of BCL for semantic segmentation. Models trained from scratch using PL-BCL and Graph Attention consistently outperform baselines across three geospatial segmentation datasets. This trend also extends to pre-trained models and settings with label scarcity.
Road network segmentation from satellite imagery remains challenging due to large geographic variation in road appearance, occlusions, and domain shifts introduced by differing resolutions and sensors. Existing models, typically trained under narrow resolution–region combinations, generalise poorly to unseen environments such as rural settings, regions with distinct road materials, or imagery from new satellite platforms, often producing broken or disconnected predictions. Adapting these models to new domains usually requires retraining or fine-tuning, which is costly and risks catastrophic forgetting. In this work, we reframe global road extraction as a continual adaptation problem rather than an architectural one. Our framework combines cross-resolution knowledge distillation across a resolution-decreasing curriculum, multi-sensor training, and topology-aware supervision, yielding a single model that generalises across 0.3-1.0 m imagery from multiple satellite platforms across continents. On publicly available benchmarks, including City-Scale and Global-Scale, our model outperforms state-of-the-art results by up to 22 F1 points and 15 APLS points, while remaining the most efficient, with 3× faster inference. Our results suggest that robust planetary-scale mapping is achievable through targeted training strategies, such as data curricula, distillation, and topology losses, rather than increasingly complex architectures.
Air pollution is a critical global challenge, causing significant mortality annually. Yet, ground-based monitoring networks remain sparse and costly, leaving unmonitored areas without reliable air quality data. Remote sensing offers a transformative alternative, enabling continuous, large-scale atmospheric monitoring across remote, unobserved regions. However, progress in deep learning for this task has been constrained by the absence of large-scale, openly available benchmark datasets. To address this gap, we introduce Sen2Air and SentX-Air, two large-scale Earth observation benchmark datasets for estimating six air pollutants (NO₂, CO, O₃, SO₂, PM₁₀, PM₂.₅) from high-resolution Sentinel-2 imagery, with SentX-Air incorporating Sentinel-5P. Spanning over 150 ground-truth monitoring stations across Australia and India, our datasets enable pollutant estimation at unprecedented 10m spatial resolution. We benchmark CNN-based, Transformer-based, and self-supervised architectures under single and multi-pollutant evaluation protocols, providing the first systematic comparison of deep learning models for pollutant estimation from multispectral satellite data. Our findings highlight the pollutant-specific impact of Sentinel-5P integration and open challenges in multi-modal fusion and cross-region generalization. Both datasets will be publicly released to accelerate EO-based air quality research.
Super-resolving coarse atmospheric fields to local PM2.5 variations is uniquely challenged by a mismatch in spatial support: while pixels represent regional averages, ground-truth observations are discrete, unaligned samples of a continuous spatial signal. To bridge this gap, we present a station-guided framework for high-resolution PM2.5 downscaling over Europe. Taking coarse CAMS atmospheric composition fields alongside heterogeneous side information (i.e., human activity, land cover, elevation, satellite aerosol observations, and wind fields) our framework jointly super-resolves (×40, ≈1km) and bias-corrects CAMS rasters, without relying on temporal sequence modelling. To address the challenge of densely supervising our multi-scale transformer network with sparse in-situ data, we introduce a time-agnostic propagation strategy that utilises spatial Gaussian blending of interpolated OpenAQ observations. Extensive qualitative and station-level evaluations across Europe demonstrate that our model recovers fine-grained spatial structures and effectively mitigates localised CAMS biases.
Despite comprising over 70% of its surface, the world's oceans are critically underobserved compared to the land surface or the atmosphere. Understanding the global ocean requires jointly observing its surface and subsurface structure, yet no standardized, high-resolution dataset couples satellite surface fields to co-located in situ depth profiles in an AI-ready format. Existing resources either consist of model-reconstructed gridded products rather than observations, cover only a single variable or basin, or operate at resolutions too coarse for mesoscale dynamics. We introduce OceanDepths, the first open, global, regridded AI-ready dataset that pairs satellite-derived sea surface temperature (SST), sea surface salinity (SSS), and sea surface height (SSH) L4 products with co-located EN4 subsurface temperature and salinity profiles, complemented by matched GLORYS12 ocean reanalysis data to support comparisons or multi-stage learning. The dataset spans 2000-2024 at 0.1° x 0.1° spatial resolution and at weekly temporal resolution, covering the entire globe's sea surface and with over 9.5 million paired profiles interpolated to 50 standardized depth levels. We provide a configurable system to split the globe in equally sized spatial patches. The 4D multivariate structure, high resolution, long temporal extent, and extreme sparsity of subsurface observations (~0.01% per depth level) make OceanDepths a challenging testbed for novel AI methods. We demonstrate subsurface state reconstruction as an example task with simple baseline models, but also envision OceanDepths to support the development of observation-based forecast methods and other related tasks.
Earth Observation (EO) text-image retrieval benchmarks reward a system for ranking the single image that produced a caption above all others. Analysts, however, issue intent queries such as “bridges still passable in the flooded area”, which may match only a small region but across many tiles, making retrieval multi-positive. We introduce SkySpector, the first multi-positive, query-centric EO retrieval benchmark: 412 queries authored or validated by EO experts, each paired with a vetted set of relevant images (avg. 4.6) in a shared pool of 972 images. Unlike a caption-derived corpus, each text is annotated against the shared pool rather than only against its seed image. Queries are stratified by semantic level (scene, region, instance) and tagged with six phenomena (entity, attribute, relation, count, layout, context), so scores can be sliced by capability. We build SkySpector on established EO datasets, inheriting public, well-sampled imagery and augmenting it with analyst-grade text and provenance-preserving relevance labels; every query keeps its source dataset and image identifiers. Evaluating 16 vision-language models shows that benchmark construction changes conclusions: model rankings that are stable across three caption benchmarks (Kendall's τ ≥ 0.72) collapse on SkySpector (τ = 0.13 vs. RSITMD), with the largest drops on relation, count, and layout queries. We posit that single-vector CLIP-style retrieval is insufficient for real-world tasks and show that even a simple multimodal-LLM reranking baseline recovers +13.9 NDCG@10, but still leaving substantial headroom on the benchmark. We will publicly release the benchmark, evaluation code, and leaderboard upon acceptance.
High quality reference data remain a critical bottleneck for crop-type mapping at any spatial and temporal scale. Operational systems such as WorldCereal aggregate labels from heterogeneous sources such as parcel registers, national databases, field surveys, and map-derived products, each with their own biases, coverage gaps and unknown label noise. Simple global rules are inadequate, since crop phenology and observation conditions vary strongly across regions and seasons. In this study, we focus on a single, operationally relevant question: whether embeddings produced through geospatial foundation models are a viable basis for cleaning the reference data. We propose a practical, locality-aware, embedding-based anomaly (EBA) detection framework that operates on the embeddings of a pretrained Earth-observation encoder. We score each labelled sample against other samples of the same crop in the same area using a pretrained embedding, flag the ones that stand out, and test whether removing or down-weighting them before training yields a better model. We establish that the flagged points are genuinely mislabelled or misplaced in two independent ways: against synthetic ground truth, the detector concentrates injected label errors 2.5–5× above chance in its flagged set (detection AUROC up to 0.84); and on real data, a model-independent test shows that removing or confidence-weighting the flagged held-out points raises measured accuracy in trained models, for both crop type and land cover. Acting on the flags then improves the WorldCereal crop-type model across five macro-regions, evaluated on a fixed held-out split under three views. We find conservative cleaning helps while over-cleaning hurts. The EBA detector approach is designed to be reproducible and extensible, and can serve as a template for cleaning large, noisy Earth observation reference datasets beyond crop mapping.
Presented in the dedicated afternoon poster session.
Recent years have seen a rapid expansion in the production of large-scale geospatial maps derived from Earth observation (EO) data, driven largely by advances in machine learning (ML) and large computing infrastructure. Although the barrier to generating such maps has dropped substantially, established best practices have yet to emerge, and design decisions made early in the pipeline can quietly propagate errors into the final product. Producing a technically sound and scientifically credible product remains challenging. Choices made at every stage are tightly coupled: preprocessing decisions shape the training signal, dataset design governs what the model can learn and how reliably its performance can be assessed, and global-scale inference introduces engineering challenges in compute and data access at scale, as well as artifact mitigation. Furthermore, uncertainty quantification and independent map validation each require dedicated methodological attention that is often underestimated. This paper presents a concise, end-to-end account of the recommended practices spanning the pipeline from satellite data to an operational map product. We organize the discussion around six interconnected themes: the EO data infrastructure landscape, data selection and preprocessing, ML dataset construction and model training, uncertainty quantification, map production and distribution, and validation. This paper is a condensed version of a longer guide that provides greater depth across all stages, accessible online at mleomaps.netlify.app.
Land surface temperature (LST) is an Essential Climate Variable (ECV) crucial for understanding land–atmosphere energy exchange and monitoring climate change, especially in the rapidly warming Arctic. Long-term satellite-based LST records, such as those derived from the Advanced Very High Resolution Radiometer (AVHRR), are essential for detecting climate trends. Data from the AVHRR sensors series cover more than four decades and thus provide an essential archive to derive trends and account for interannual variability. However, the coarse spatial resolution of AVHRR Global Area Coverage (GAC) data limits their utility for analyzing fine-scale permafrost dynamics and other surface processes in the Arctic. This paper presents a new 42-year pan-Arctic LST dataset, downscaled from AVHRR GAC to 1 km using a guided super-resolution algorithm based on a deep anisotropic diffusion model. The model is trained on MODIS LST data, using coarsened inputs and native-resolution outputs, guided by high-resolution land-cover, digital elevation, and vegetation-height maps. The resulting dataset provides twice-daily, 1 km LST observations for the entire pan-Arctic region spanning more than four decades. The dataset is validated against in situ observations from ground stations across the United States, the Arctic, and Europe, demonstrating its ability to reproduce local-scale temperature variability. In addition, the downscaled product is intercompared with an independent 1 km AVHRR LST product produced by EUMETSAT, providing further assessment of its spatial and temporal consistency. This enhanced dataset enables improved modelling of permafrost, reconstruction of near-surface air temperature, and assessment of the surface mass balance of the Greenland Ice Sheet. Additionally, it supports climate monitoring efforts in the pre-MODIS era and offers a framework adaptable to future satellite missions for thermal infrared observations and climate data record continuity.
Rapid disaster response can benefit from detecting changes directly onboard Earth observation satellites, yet the datasets available to train and evaluate such models are scattered across disaster types and sensors, and none support true continuous change detection. We present the current state of RestlessEarth, a multi-temporal Sentinel-2 dataset that unifies three disaster types - floods, wildfires, and landslides - under a single source with a worldwide coverage targeting area of 850,409 km². We reuse only the disaster locations, segmentation masks, and event date from a set of public datasets, and re-acquire Sentinel-2 timeseries using an automated, cloud-aware acquisition pipeline. The pipeline selects a sequence of the least obstructed scenes spanning 12 months before and 2 months after each event, enabling continuous change detection and onboard-style data simulation. Masks are refined by human annotators following a framework designed around the specific challenges of multi-spectral, temporal satellite imagery. Our approach supports advanced sensors, using SWIR bands, but also low-cost SmallSat cameras restricted to RGB and NIR bands.
The increased frequency and severity of natural disasters, intensified by climate change, lead to significant humanitarian and economic consequences. Rapid and reliable disaster mapping is essential for coordinated response efforts, resource allocation, and the protection of lives. Recent advances in artificial intelligence and the growing volume of remote sensing data offer new opportunities to improve disaster response capabilities. However, existing datasets and models show gaps in multimodal fusion, temporal resolution, global coverage, dataset scale, and open access. We address these limitations by introducing ImpactMesh, large-scale disaster mapping datasets for floods and wildfires that integrate multitemporal Sentinel-1 and Sentinel-2 imagery with elevation data and high-quality annotations from the Copernicus Emergency Management Service. The datasets cover over 400 flood and wildfire events across diverse geographic regions, with four temporal observations per event to support pre- and post-disaster comparison. We provide preliminary results using two geospatial foundation models and publicly release the datasets under a permissive license to advance research and applications.
Vision Transformers (ViT) dominate computer vision. However, their reliance on rigid patch projectors hinders transfer to Earth Observation (EO), where inputs vary widely in modality, scale, and resolution. We introduce UniverSat, a ViT-style backbone built around a Universal Patch Encoder that maps patches from arbitrary spatial, spectral, and temporal resolutions, and from both optical and non-optical sensors, into a shared embedding space with a single set of weights. This enables training one model on heterogeneous multimodal corpora in self-supervision, yielding robust sensor-agnostic spatial features. We validate this approach with strong results across classification and segmentation on standard EO benchmarks from GeoBench, PANGEABench, and SpectralEarth. The code and model are open-source.
Vision-language models (VLMs) offer transformative potential for zero-shot remote sensing, but adapting them to synthetic aperture radar (SAR) is fundamentally constrained by scarce aligned image-text data and a deep physical domain gap. Current approaches employ aggressive fine-tuning of both image and text encoders, distorting pretrained feature spaces and demanding prohibitive data scales. We propose a Frozen Expert framework that bridges a frozen, physically grounded SAR encoder (CROMA) with a frozen CLIP text encoder via a structure-preserving projection bridge, requiring no encoder retraining. Procrustes analysis confirms that our bridge preserves CROMA's manifold geometry better than any fine-tuning strategy. Empirically, our approach achieves state-of-the-art zero-shot SAR scene classification using up to 8× less training data than leading baselines, with gains across DFC2020 (+10.3% R@1, +15.5% mAP), EuroSAT-SAR (+15.7% R@1, +13.1% mAP), and BigEarthNet-MM (+14.7% R@1).
High-resolution (HR) satellite images enable highly detailed, fine-grained mapping but are often costly and sparsely available. Meanwhile, low-resolution (LR) data sources such as Sentinel-1 and Sentinel-2 provide free, global coverage, albeit at coarser scales. This challenge thus raises the question of whether knowledge from HR imagery can be effectively transferred to models operating only on LR inputs at inference time to generate HR outputs. In this work, we investigate knowledge distillation (KD) as a privileged learning (PL) strategy for HR flood mapping, using 3 m resolution PlanetScope imagery as a privileged modality and Sentinel-1 and Sentinel-2 imagery as LR inputs. Our results show that KD methods consistently outperform non-PL methods, with logit-based KD achieving the best performance (0.621 mIoU, 0.755 F1) compared to the strongest non-PL baseline (0.592 mIoU, 0.728 F1). These findings demonstrate the potential of KD as a practical strategy for leveraging HR images during training while relying only on low-cost LR data at deployment. This ongoing work represents a first step towards a broader comparative evaluation of PL for remote sensing applications.
Applying deep learning models to satellite imagery inside geographic information systems (GIS) remains high-friction for remote sensing practitioners. Models arrive in incompatible formats and target different compute environments, from local workstations to serverless cloud services. As a result, every evaluation demands custom deployment, tiling, and georeferencing code before a single prediction reaches the analyst's map. This friction discourages systematic comparison in a domain where model choice directly affects operational outcomes such as field delineation, crop monitoring, and disaster response. We present Anaximander, an open-source system that unifies both dimensions behind one interactive interface. The system's backend is an inference server that loads models from multiple commonly-used sources and serves them on any accessible compute backend. The server provides session management, model caching, multi-device replication, batching, and streamed per-tile confidence. The backend is paired with a QGIS plugin that drives tiling, result reassembly, georeferencing, and real-time per-tile status visualization. An additional user-interface path injects layer legends as prompts into vision-language models. We demonstrate the system in a code-free side-by-side comparison of three heterogeneous models on an agricultural field delineation task: gpt-image-1 via an Azure API, Segment Anything Model 3 (SAM 3) on a remote GPU, and DelineateAnything on a local CPU. The inference backend and protocol are open-source.
In this paper, we present LCMonitor, a work in progress to build a global catalog of land cover changes with quarterly updates. Existing change detection methods are ill-suited for real-world global deployment: they often depend on expensive commercial high-resolution images, and are limited to comparing pairs of images. Thus, we annotate a new dataset of 1.3K points exhibiting change and 40K no-change points; we label each changed point with source and destination land cover types, along with precise pre-change and post-change dates. We pair these points with time series of Sentinel-2 images that incorporate four years of historical context, and develop a model capable of leveraging that context to robustly distinguish short-term variations from long-term changes. We compare LCMonitor to annual land cover map and embedding baselines on a held-out test dataset. Ongoing work includes additional annotation and modeling exploration to further improve accuracy; following this, we will deploy the first version of the catalog.
Wildfire burned-area mapping is essential for damage assessment, emissions modeling, and understanding fire–climate interactions across diverse ecological regions. Recent geospatial foundation models provide strong general-purpose representations for satellite imagery, yet there is still no clear understanding of how to efficiently adapt these models for downstream Earth observation tasks, particularly under geographic and temporal domain shift. This study evaluates three state-of-the-art Geospatial Foundation Models (GFMs) - Terramind, DINOv3, and Prithvi-v2 - for burned-area mapping across the United States and Canada using Sentinel-2 data. Leveraging 3,820 wildfire events from 2017–2023, we conduct spatial and temporal generalization tests across diverse biomes. We systematically compare full fine-tuning, decoder-only fine-tuning, and Low-Rank Adaptation (LoRA) for adapting each model. Across all experiments, LoRA provides the strongest cross-domain generalization while updating less than 1% of parameters, demonstrating a favorable trade-off between accuracy and efficiency. Prithvi-v2 with LoRA achieves the highest overall accuracy and the larger improvement compare to full fine-tuning. These findings indicate that geospatial foundation models, when adapted using lightweight parameter-efficient methods such as LoRA, offer a robust and scalable solution for large-scale burned-area mapping.
Cross-modal ship re-identification (ReID) is challenging and underserved: synthetic aperture radar (SAR) and multispectral optical imagery observe different physical signals with different failure modes, and identity matching demands a precise, leakage-free evaluation protocol that existing ship datasets rarely provide. We introduce ShipCross-ID, a benchmark that links FUSAR SAR vessel images with Sentinel-2 multispectral vessel crops through AIS-associated identities. ShipCross-ID defines fixed train, validation, and test splits over 228 eligible identities, a shared mixed gallery, three directional query protocols, with MS denoting multispectral (MS→SAR, SAR→MS, MS→Mixed), and explicit open-set queries with no gallery match. We provide a DINOv3 11-channel baseline, selected by validation mAP, as a reproducible reference. The baseline reaches ALL retrieval mAP 30.0 at Rank-1 29.2 but collapses to Rank-1 14.3 on the pure cross-modal directions, showing that SAR/multispectral ship ReID is feasible yet far from solved.
Cloud occlusion is a persistent obstacle for optical satellite imagery, yet generative models for satellite data rarely offer explicit, controllable handling of clouds. We address this gap by constructing a new RGB satellite image dataset for the Guinea Forest of West Africa region, in which each image is paired with its GPS coordinate, acquisition date, and a three-class cloud mask. Building on this dataset, we adapt the JiT diffusion model to synthesize RGB satellite images directly from these metadata alone (GPS, acquisition date and cloud mask) without any input image, using an auxiliary cloud loss and cloud conditioned classifier-free guidance. Evaluated with both reconstruction quality and cloud segmentation metrics, and show our finetuned model generates high-fidelity satellite images with more accurate clouds. By varying the cloud mask, the same model performs cloud controlled generation for both cloud removal and cloud synthesis tasks, enabling flexible and controllable cloudy image synthesis.
Satellite imagery enables global environmental monitoring, yet the annotations available for it - coarse land cover classes, noisy captions, or proxy labels - pale in comparison to the fine-grained semantics captured by a ground-level photograph. This paper presents GLUE (Geometrically-Lifted Unified Earth Embeddings), a method that distils fine-grained visual knowledge from ~205,000 geo-referenced ground photographs into dense satellite embeddings via multi-view geometric reconstruction. The result is an open-vocabulary system: any phrase in any of 109 languages produces a dense semantic heatmap at 10 m resolution. Without ever seeing a category label, GLUE achieves 50.7% zero-shot 8-class accuracy (vs. 45.7% for the best multispectral baseline) and a caption agreement of 0.912 against ground-level observations (vs. 0.549). A learned query adapter enables bare-text querying across languages, generalising to held-out Chinese with cross-lingual consistency comparable to languages seen during training.
Thanks to our reviewers, and in particular to the emergency reviewers who contributed to this crucial and laborious process. Each paper was assigned three reviewers — two for short papers — and handled by an area chair.
This year we did not blanket desk-reject submissions — nor did we remove reviews — that were suspected of LLM writing. We assessed every case independently, and based our decisions on scientific soundness, factual accuracy and validity of the results. This took more time, but allowed for flexibility in how we handled borderline cases. Next year we plan to improve by releasing appropriate guidelines in the main call for papers. In line with the ECCV rules, the use of LLMs is not in itself forbidden, but authors and reviewers are fully responsible for the content they contribute, and inconsistencies are treated the same way regardless of whether they are human-generated or not.
We are grateful to all the authors who submitted to TerraBytes. The diversity of topics in this year’s edition is exciting, and we are glad to see how quickly the workshop has become an attractive venue at the crossroads of methods, data and applications of machine learning and computer vision for Earth observation.
The call closed on 25 June 2026 and peer review is complete — decisions were sent on 24 July 2026 and 36 papers were accepted. The scope and topics below are kept for reference.
TerraBytes II is held in conjunction with ECCV 2026 in Malmö, Sweden. TerraBytes brings the Earth observation, machine learning, and computer vision communities together to discuss global datasets and models for our planet — covering every step from downlinked satellite data to training paradigms and downstream applications.
We welcome submissions on (but not limited to):
Submission to the TerraBytes workshop was double-blind. We accepted short papers (8 pages excluding references) and full-length papers (14 pages excluding references), in the standard ECCV 2026 paper format.
Short papers can describe work in progress, opinion pieces, or datasets and research papers that have been published in another venue (conference or journal) within the last 6 months on topics relevant to TerraBytes. For already-published work, please submit a summarised short paper — journal papers are particularly welcome, as the workshop is a great venue to bring them to a wider audience.
All submissions go through OpenReview. Each paper is reviewed by at least two non-conflicting reviewers. All accepted papers are presented in a dedicated poster session; full-length papers and selected short papers will also be given spotlight presentations throughout the day.
Supplementary material may be uploaded alongside the paper via OpenReview and will be published with accepted submissions. Reviewers may consult supplementary material but are under no obligation to do so — the paper itself must be self-contained and stand on its own merits during review.
It is expected that at least one author of each accepted paper will register for the workshop and present the paper in person. Online presentations will be considered for presenters who cannot travel (e.g. visa or personal constraints).
All deadlines are 23:59 Anywhere on Earth.
Accepted full-length papers can be published in an archival proceedings volume of the ECCV workshops (opt-in). Authors who prefer their paper to remain non-archival — for example to preserve future submission options — may choose to submit short papers instead.
Second edition of TerraBytes, returning to build on the momentum of 2025. The main ECCV conference runs 8–13 September at the Malmö Arena and Malmömässan. 36 accepted papers, three keynotes and a panel on global models.
24 accepted papers, three keynotes (NASA IMPACT, ESA Ground Segment, UAF) and a panel on Global Geospatial Models. Recordings and proceedings available.
Workshop social
After the closing remarks, join the TerraBytes community for an informal evening at a separate venue — a chance to carry the day’s conversations on over food and drinks.
Starts 18:30. Exact time and venue are still being confirmed; the Luma page carries the address and any changes. Sign-up is required so we can plan numbers.