What is Granite Geospatial UKI?
Granite Geospatial UKI is an IBM Granite geospatial foundation model for analyzing satellite imagery over the United Kingdom and Ireland. Its purpose is to provide a reusable visual representation of Earth-observation data. Rather than training a separate model from scratch for every regional remote-sensing task, a practitioner can use the pretrained encoder as a starting point and fine-tune it for a downstream application.
The model is built around a Vision Transformer, a neural-network architecture that processes an image as a collection of smaller patches. It follows the Prithvi-EO design and uses masked-autoencoder pretraining. During this type of self-supervised training, portions of the input imagery are hidden and the model learns to reconstruct or represent the missing information. This allows it to learn patterns in satellite data without requiring every training image to have a manually created label.
Granite Geospatial UKI is not a language model and should not be evaluated like a chatbot. It does not produce written answers, code, audio, video, or generated pictures. Its practical output is a learned representation that can support a downstream computer-vision model, such as a segmentation system that assigns a class to each image pixel.
Provider and position in IBM’s catalog
The model is provided by IBM through the IBM Granite public model ecosystem and is hosted as an open-weight repository on Hugging Face. It fits the specialized research and geospatial side of IBM’s broader AI portfolio rather than the consumer-chat or general-purpose assistant category.
IBM’s watsonx portfolio includes enterprise AI development, model deployment, governance, retrieval, agents, and access to IBM Granite and other models. However, the supplied research does not list Granite Geospatial UKI as a hosted watsonx inference endpoint. The model’s documented distribution is an open-weight repository, and its usage examples rely on TerraTorch. This distinction matters: access to the broader IBM watsonx platform does not by itself establish a managed API, a chat interface, or a product-specific service for this model.
Satellite data and architecture
The model combines two kinds of Earth-observation data. Harmonized Landsat and Sentinel-2, commonly abbreviated HLS, provides multispectral optical information. Sentinel-1 provides synthetic-aperture radar, or SAR, data. Radar is useful because it captures information differently from optical sensors and can remain informative in conditions where clouds or lighting make optical imagery less useful.
The documented input bands are Blue, Green, Red, Narrow NIR, SWIR 1, SWIR 2, VV, and VH. The first six are multispectral optical bands from HLS, while VV and VH are Sentinel-1 radar polarization channels. In practical terms, the model can examine visible light, near-infrared and short-wave infrared characteristics alongside radar responses.
Initial pretraining used HLS data covering the continental United States. IBM and its research collaborators then performed additional pretraining using approximately 16,000 HLS L30 and Sentinel-1 images covering the United Kingdom and Ireland. That regional stage is the key reason to consider this model for UK and Ireland workflows: its representation is adapted to the geographic area rather than being only a generic global remote-sensing encoder.
The model card documents use with TerraTorch, a geospatial deep-learning framework. The repository also includes a downstream configuration for fine-tuning. The supplied configuration research refers to a nine-band downstream setup that includes a cloud band, but the model’s documented core band list contains the eight optical and radar channels described above. Users should therefore check the repository configuration and preprocessing requirements before assuming that every downstream checkpoint accepts exactly the same inputs.
What is it designed to do?
Granite Geospatial UKI is primarily designed for remote-sensing feature extraction and transfer learning. A feature extractor converts raw imagery into numerical representations that capture spatial and spectral patterns. A downstream task-specific model can then use those representations for classification, segmentation, detection, or related analysis.
The clearest documented application is regional flood segmentation. In a segmentation task, the system predicts a category for each pixel or image region, such as flooded area and non-flooded area. A related flood-detection checkpoint adds a fully convolutional decoder and a two-class segmentation head. That checkpoint is a downstream model rather than the same entity as the base Granite Geospatial UKI model, so its results and architecture should not be treated as specifications of the base encoder.
Other reasonable uses supported by the model’s design include:
- Extracting representations from multispectral and SAR imagery for research experiments.
- Fine-tuning a regional classifier or segmentation model.
- Studying flood extent and other environmental patterns over the UK and Ireland.
- Building Earth-observation pipelines that combine optical and radar satellite observations.
- Using a pretrained regional encoder when labeled data is limited and training a full vision model from scratch would be inefficient.
Reported results and important limitations
The research paper reports whole-test-set mean intersection over union, or mIoU, of 0.786 and F1 of 0.938 for the related flood-detection setup. These are results for the documented downstream flood-segmentation experiment, not a universal score for Granite Geospatial UKI on every task. They should not be interpreted as a general accuracy guarantee for new regions, sensors, preprocessing pipelines, or label definitions.
The paper identifies a specific data-alignment risk: mismatched Sentinel-1 and Sentinel-2 swaths can produce artifacts. Predictions should therefore be restricted to areas where the relevant satellite swaths overlap. This is a practical limitation rather than a minor implementation detail. A pipeline that combines imagery from different acquisitions without checking spatial coverage and alignment may generate misleading segmentation results.
The model is also region-specific. Its UK and Ireland focus can be an advantage for those areas, but it does not establish equivalent performance in other countries. Teams working in a different geographic region may prefer a broader geospatial foundation model or may need additional regional pretraining and fine-tuning.
Finally, the base model is not a complete end-user application. It does not automatically provide a finished flood map, dashboard, labeling workflow, monitoring service, or hosted prediction endpoint. Those functions require additional preprocessing, task-specific model components, compute infrastructure, and validation.
Inputs, outputs, and unsupported expectations
The model accepts satellite imagery rather than text prompts. Its documented inputs include HLS multispectral channels and Sentinel-1 VV and VH SAR channels. It is therefore multimodal in the remote-sensing sense of combining optical and radar data, but it is not a general multimodal assistant that accepts arbitrary images, audio, video, and text in a conversational interface.
Granite Geospatial UKI does not have a documented text output, image-generation output, audio output, video output, or music output. Its useful result is an internal visual representation or, when paired with a downstream decoder, a task-specific prediction such as a segmentation mask. No context-window limit or maximum generated-token limit is applicable in the language-model sense, and no official output-token limit is listed.
There is no verified support for function calling, external tools, web search, streaming responses, JSON-mode generation, or chat. It is likewise not a coding model and should not be selected for software development or code completion. Any code used with it is integration or training code around the vision model, not generated programming output from the model itself.
Pricing and access
The supplied research lists no official hosted API pricing for Granite Geospatial UKI. It is described as an open-weight research model rather than a metered inference product. The repository is licensed under Apache-2.0, and IBM’s public-repository disclosure states that the code is provided as an open-source project without an obligation for IBM to maintain or support it.
Open weights can reduce licensing barriers and give a team more control over deployment, but they do not make operation cost-free. Users may need storage for satellite imagery, preprocessing tools, compatible hardware, model-serving infrastructure, and engineering time for fine-tuning and validation. Actual cost will depend on image volume, resolution, batch size, hardware, and the complexity of the downstream task.
Because no hosted endpoint is listed, readers should not assume that a simple API key and per-request billing path are available. The practical access route is to obtain the repository and configure an appropriate geospatial machine-learning environment, following IBM’s documented TerraTorch example.
Speed, cost, and reasoning trade-offs
This model trades general-purpose convenience for specialization. Compared with a hosted multimodal assistant, it is less convenient because it requires imagery preparation, model loading, downstream task design, and infrastructure. Compared with training a geospatial model from the beginning, it can reduce the amount of task-specific training needed by supplying a pretrained regional representation.
There is no provider-published speed benchmark or hardware requirement in the supplied research. It would therefore be misleading to promise a particular throughput or latency. In practice, performance will depend on the model implementation, image dimensions, batch size, accelerator, data pipeline, and whether the base encoder is frozen or fine-tuned.
The model also does not perform reasoning in the language-model sense. It learns visual and geospatial patterns, but it does not explain a result in natural language or independently plan a multi-step investigation. If a workflow needs a written interpretation, an analyst or a separate language model would need to process the model’s outputs.
When to choose Granite Geospatial UKI
Choose Granite Geospatial UKI when the project has a strong connection to the United Kingdom or Ireland, uses the documented optical and radar satellite inputs, and needs a foundation for a custom Earth-observation task. It is especially suitable when a team wants open-weight access, regional pretraining, and the ability to fine-tune a model for flood segmentation or similar remote-sensing work.
It is a good candidate for research groups, environmental-monitoring teams, and developers who can manage a Python-based geospatial machine-learning workflow. It is also attractive when keeping the model and data pipeline under the team’s control is more important than using a turnkey hosted service.
Another option may be more appropriate when the goal is text generation, conversational analysis, software development, direct image generation, speech, or general image understanding. A different geospatial model may also be preferable for global coverage, a different sensor combination, a managed inference API, or a finished operational application. Even for UK and Ireland projects, teams should compare against alternative encoders if their imagery does not contain the required bands or if their labels and geographic coverage differ substantially from the training and evaluation setup.
Bottom line
Granite Geospatial UKI is best understood as a specialized, open-weight regional encoder for satellite-image research and downstream Earth-observation tasks. Its distinguishing value is the combination of UK and Ireland-focused pretraining with HLS multispectral and Sentinel-1 radar data. Its main limitations are equally clear: it is not a chatbot, it has no listed hosted API price, it requires a separate downstream task design, and its results depend heavily on correct sensor alignment and regional validation.