SatDINO: Self-Supervised Satellite Representation

• SatDINO is a DINO-based self-supervised framework for remote sensing that uses Vision Transformers and tailored GSD encoding to address scale variations in satellite imagery.
• It employs a teacher-student architecture with adaptive multi-crop sampling and a dedicated GSD token to learn robust, label-free representations across heterogeneous Earth-observation datasets.
• Empirical results demonstrate that SatDINO outperforms MAE-based baselines in classification and segmentation tasks and serves effectively as a frozen encoder for various downstream applications.

SatDINO denotes a DINO-based approach to remote-sensing representation learning whose primary formulation is a self-supervised Vision Transformer pretraining framework tailored to satellite imagery, with explicit mechanisms for scale robustness and downstream transfer across heterogeneous Earth-observation datasets (Straka et al., 29 Aug 2025). In a broader usage that appears in later work, “SatDINO” also denotes the practice of treating a satellite-pretrained DINO backbone as a frozen general-purpose encoder, or “world-model,” for downstream Earth-observation tasks such as probabilistic rainfall nowcasting (Filho et al., 14 Nov 2025). Across these usages, the unifying idea is that DINO-style self-distillation can serve as a foundation for satellite imagery semantics, scale variation, and label-scarce transfer.

1. Terminological scope and historical positioning

In its most specific sense, SatDINO is the model introduced in “SatDINO: A Deep Dive into Self-Supervised Pretraining for Remote Sensing,” where it is presented as a DINO-style self-supervised pretraining framework adapted to satellite imagery and evaluated against masked autoencoder baselines such as SatMAE, Scale-MAE, and Cross-Scale MAE (Straka et al., 29 Aug 2025). That work positions SatDINO as a remote-sensing alternative to the dominant masked image modeling family, arguing that DINO’s multi-crop invariance objective is naturally compatible with the multi-scale character of satellite imagery.

A second, broader sense appears in work on probabilistic rainfall nowcasting, where “SatDINO” refers to using a large DINO-style satellite foundation model—specifically DINOv3-SAT493M ViT-L/16—as a frozen encoder for Earth-observation tasks (Filho et al., 14 Nov 2025). Under this usage, SatDINO is less a single training recipe than a deployment pattern: a satellite-specific DINO backbone, kept fixed, with thin task-specific heads or projectors trained on top.

A common misconception is that SatDINO is merely “DINO trained on satellite data.” The dedicated SatDINO pretraining paper explicitly adds remote-sensing-specific mechanisms, most notably a new way to incorporate ground sample distance (GSD) encoding and an adaptive or uniform local-view sampling strategy, and it evaluates these components through targeted ablations (Straka et al., 29 Aug 2025). Conversely, the broader encoder-centric usage shows that SatDINO is also a systems concept: a satellite-pretrained DINO backbone can act as a reusable prior for spatiotemporal Earth-observation pipelines (Filho et al., 14 Nov 2025).

2. Core pretraining framework and objective

The canonical SatDINO formulation retains the standard DINO teacher–student architecture with Vision Transformers as backbones, but specializes the training protocol to remote sensing (Straka et al., 29 Aug 2025). Two networks are used: a student network $f_s(\cdot;\theta_s)$, trained by gradient descent, and a teacher network $f_t(\cdot;\theta_t)$, updated through exponential moving average. For each image $x$, the model generates two global views covering $25\%$–$100\%$ of the image and $n_l$ local views covering $5\%$–$25\%$, resized to fixed resolutions and processed through projection heads that output distributions over a prototype space.

The DINO objective is a cross-view KL divergence. In simplified form, with global views $\mathcal{V}_t$ for the teacher and all views $\mathcal{V}_s$ for the student,

$f_t(\cdot;\theta_t)$0

with

$f_t(\cdot;\theta_t)$1

Teacher logits are centered to avoid collapse, and the model follows the standard DINO multi-crop regime rather than masked reconstruction (Straka et al., 29 Aug 2025).

SatDINO uses standard ViT backbones, but augments the tokenization scheme with an additional GSD token alongside the class token. The final GSD token embedding is passed through a linear regression head to predict ground sample distance directly from image content. This produces the combined objective

$f_t(\cdot;\theta_t)$2

where

$f_t(\cdot;\theta_t)$3

The design goal is to encode scale information without requiring GSD metadata at inference time (Straka et al., 29 Aug 2025).

Three model variants are reported.

Variant	Parameters	GFLOPs
SatDINO-Small$f_t(\cdot;\theta_t)$4	21.59M	8.54
SatDINO-Small$f_t(\cdot;\theta_t)$5	21.37M	33.56
SatDINO-Base	85.65M	33.90

These variants preserve the standard ViT patch-token pipeline while exposing a trade-off between patch granularity and compute. SatDINO-Small$f_t(\cdot;\theta_t)$6 uses finer $f_t(\cdot;\theta_t)$7 patching and achieves the strongest overall reported linear-probing result on the pretraining dataset, while SatDINO-Base increases capacity but does not uniformly dominate downstream transfer (Straka et al., 29 Aug 2025).

3. Scale awareness: GSD encoding and adaptive view sampling

SatDINO’s most distinctive contribution is its treatment of scale, which is central in remote sensing because object appearance varies with GSD, platform, and spatial resolution (Straka et al., 29 Aug 2025). The paper frames GSD as affecting object size in pixels, texture statistics, and cross-dataset transfer, and contrasts its approach with MAE-based methods that inject GSD as metadata into positional encodings.

The GSD mechanism is optional at the framework level but present in the recommended SatDINO configuration. A randomly initialized GSD token is appended to the ViT sequence, and only the student’s GSD prediction is penalized. Ablations over $f_t(\cdot;\theta_t)$8 show that smaller $f_t(\cdot;\theta_t)$9 can slightly improve fMoW linear probing, whereas $x$0 improves downstream classification robustness, especially across GSD shifts; the final choice is $x$1 (Straka et al., 29 Aug 2025).

The second scale-aware modification is adaptive or uniform local-view sampling. Instead of sampling every local crop independently from the same interval, SatDINO divides the local crop-scale range into equal subranges and assigns each local view to one subrange. If the scale interval is partitioned into $x$2, then each local view samples

$x$3

This ensures that the set of local crops for a single image spans a wider range of effective scales after resizing, approximating variation in GSD. The authors compare this strategy with fully random sampling and with DINO-MC-style variable local resolutions, concluding that uniform local scale sampling improves fMoW metrics and, when combined with GSD encoding, improves both pretraining and downstream metrics (Straka et al., 29 Aug 2025).

The final SatDINO configuration uses default DINO augmentations, no temporal augmentations, $x$4–$x$5 local views, local scale range $x$6–$x$7, global scale range $x$8–$x$9, and uniform local sampling. The paper also reports that the default crop aspect-ratio range $25\%$0 is preferable to strict square crops $25\%$1, with square crops reducing kNN performance (Straka et al., 29 Aug 2025). This suggests that some geometric distortion induced by non-square crops and resizing is beneficial even in overhead imagery.

4. Pretraining corpus, transfer benchmarks, and empirical behavior

SatDINO is pretrained on fMoW-RGB, using $25\%$2 RGB images spanning $25\%$3 categories, with a GSD range of $25\%$4–$25\%$5 m/pixel and $25\%$6 time series (Straka et al., 29 Aug 2025). Labels are not used for self-supervision, although GSD metadata is used to supervise the GSD head. For SatDINO-Small, the optimizer is AdamW with cosine decay and warm-up; SatDINO-Small$25\%$7 is trained for $25\%$8 epochs and SatDINO-Small$25\%$9 for $100\%$0 epochs. Batch size is $100\%$1 per GPU on $100\%$2 GPUs for the patch-size-16 model and $100\%$3 per GPU on $100\%$4 GPUs for the patch-size-8 model.

On fMoW-RGB linear probing, SatDINO outperforms the MAE-based baselines listed in the paper despite using smaller architectures. Reported top-1 accuracies are $100\%$5 for SatDINO-Small$100\%$6, $100\%$7 for SatDINO-Small$100\%$8, and $100\%$9 for SatDINO-Base, compared with $n_l$0 for SatMAE and $n_l$1 for a Cross-MAE ViT-Base variant (Straka et al., 29 Aug 2025). Under full fine-tuning, SatDINO remains competitive rather than uniformly dominant: SatDINO-Small$n_l$2 reaches $n_l$3 top-1 and $n_l$4 top-5, comparable to Scale-MAE’s $n_l$5 top-1 and $n_l$6 top-5.

The transfer evaluation is deliberately broad. For classification, the paper studies EuroSAT, RESISC45, UC Merced, WHU-RS19, RS-C11, and SIRI-WHU, covering GSDs from $n_l$7 m to $n_l$8 m and image sizes from $n_l$9 to $5\%$0 (Straka et al., 29 Aug 2025). In average kNN accuracy across scales, SatDINO-Small$5\%$1 reports $5\%$2 on EuroSAT, $5\%$3 on RESISC45, $5\%$4 on UC Merced, $5\%$5 on WHU-RS19, $5\%$6 on RS-C11, and $5\%$7 on SIRI-WHU; these values exceed the corresponding Scale-MAE numbers on every listed dataset. Under downstream fine-tuning, all three SatDINO variants obtain very high classification accuracy, including $5\%$8 on EuroSAT for SatDINO-Base and $5\%$9 on UC Merced for SatDINO-Small$25\%$0 (Straka et al., 29 Aug 2025).

For semantic segmentation, the paper uses SatDINO as a ViT backbone inside UperNet and evaluates on Potsdam, Vaihingen, and LoveDA at $25\%$1 and $25\%$2 input resolutions (Straka et al., 29 Aug 2025). The most notable result is that SatDINO-Small$25\%$3 surpasses Scale-MAE at $25\%$4 on Potsdam ($25\%$5 versus $25\%$6) and Vaihingen ($25\%$7 versus $25\%$8), but Scale-MAE retains an advantage on several $25\%$9 settings, especially on LoveDA and Vaihingen. The paper interprets this as evidence that DINO-style pretraining is particularly strong for embedding-oriented tasks such as kNN, linear probing, and classification, while MAE-based ViT-Large models still hold an edge in some high-resolution dense prediction regimes.

5. SatDINO as a frozen foundation encoder for Earth observation

A later usage of the term shifts the emphasis from pretraining to deployment. In “DINOv3 as a Frozen Encoder for CRPS-Oriented Probabilistic Rainfall Nowcasting,” SatDINO refers to DINOv3-SAT493M ViT-L/16, a DINOv3 backbone trained on approximately $\mathcal{V}_t$0M multispectral satellite patches and used as a fixed encoder for Weather4Cast 2025 rainfall nowcasting (Filho et al., 14 Nov 2025). The encoder is treated as a frozen satellite “world-model,” and only a lightweight spatiotemporal projector and probabilistic head are trained.

The downstream input differs from the original pretraining modality: Weather4Cast provides $\mathcal{V}_t$1 bands over $\mathcal{V}_t$2 satellite frames, with each frame originally at $\mathcal{V}_t$3. The system first resizes each frame to $\mathcal{V}_t$4 and applies a learned $\mathcal{V}_t$5 channel projection to form pseudo-RGB inputs compatible with DINOv3-SAT. The frozen encoder produces $\mathcal{V}_t$6 patch tokens per frame, each $\mathcal{V}_t$7-dimensional, yielding

$\mathcal{V}_t$8

These tokens are processed by a V-JEPA-style asymmetric encoder–decoder ViT, temporally expanded from $\mathcal{V}_t$9 to $\mathcal{V}_s$0 forecast steps, and then collapsed by a 3D convolution into logits over rainfall bins (Filho et al., 14 Nov 2025).

The probabilistic output is a discrete empirical CDF over rainfall, trained with the Rank Probability Score as a discrete CRPS surrogate. In the best configuration, the head uses $\mathcal{V}_s$1 mm/hr and $\mathcal{V}_s$2 bins. On the Weather4Cast 2025 benchmark, the DINO-VJEPA system achieves CRPS $\mathcal{V}_s$3, which the authors state is an approximately $\mathcal{V}_s$4 effectiveness gain against the best 3D-UNET baseline at $\mathcal{V}_s$5 (Filho et al., 14 Nov 2025). Fine discretization is reported as critical: a coarser $\mathcal{V}_s$6 version of DINO-VJEPA yields CRPS $\mathcal{V}_s$7.

This broader formulation shows that SatDINO can function as an encoder prior rather than as a complete task model. The frozen-backbone pattern is especially attractive in settings with sparse labels, domain alignment to satellite imagery, and a need for computational efficiency, because no backpropagation through the large DINOv3 backbone is required (Filho et al., 14 Nov 2025).

6. Related DINO-based extensions, misconceptions, and limitations

The literature around SatDINO also includes adjacent architectures that are not themselves SatDINO but explicitly serve as templates for a broader “SatDINO family.” FS-DINO is a few-shot semantic segmentation framework built on a frozen DINOv2 ViT-B/16 encoder, a bottleneck adapter, a meta-visual prompt generator, and a SAM-derived mask decoder; although it is trained and evaluated on natural-image benchmarks rather than satellite data, the paper explicitly presents it as a blueprint for a “SatDINO”-style few-shot segmenter in specialized domains (Zhuo et al., 22 Apr 2025). Its proposed transfer path is clear: replace DINOv2 with a satellite-pretrained ViT, distill from a domain-specific teacher, and retain the support–query prompting and 4D correlation machinery.

SemDINO occupies a similar position for temporal remote-sensing tasks. It is an end-to-end semantic change detection network that combines a trainable ResNet50+FPN branch with frozen DINOv3 features fused by gated pyramid fusion, followed by a multi-scale temporal bidirectional transformer interaction module and several change-purification components (Tong et al., 8 Jun 2026). The paper is not titled SatDINO, but it explicitly frames its design principles as informative for a broader DINO-based remote-sensing family. On Landsat-SCD, SemDINO reports OA $\mathcal{V}_s$8, mIoU $\mathcal{V}_s$9, SeK $f_t(\cdot;\theta_t)$00, and $f_t(\cdot;\theta_t)$01, improving over BT-SCD on all four listed metrics; on LEVIR-CD it reports F1 $f_t(\cdot;\theta_t)$02 and IoU $f_t(\cdot;\theta_t)$03 (Tong et al., 8 Jun 2026). This suggests that frozen DINO semantics can be integrated effectively with multi-scale temporal alignment and change-specific heads.

Several limitations recur across the SatDINO literature. In the original pretraining paper, temporal augmentations on fMoW sequences are explored but omitted from the final configuration because they produce mixed or negative downstream effects, and scaling DINO to ViT-Large shows signs of overfitting on fMoW (Straka et al., 29 Aug 2025). The same paper also shows that MAE-based methods retain an advantage in some high-resolution dense prediction settings. In the rainfall-nowcasting formulation, very fine discretization improves CRPS but creates instability, making early stopping mandatory; the study is also limited to a single benchmark track (Filho et al., 14 Nov 2025). A plausible implication is that SatDINO is strongest, at present, as a transferable representation substrate whose effectiveness depends on careful adaptation of scale handling, decoder design, and task-specific objectives rather than as a universally dominant end-to-end architecture across all remote-sensing problems.