InstaGeo: Compute-Efficient Geospatial Machine Learning from Data to Deployment

Ibrahim Salihu Yusuf, Iffanice Houndayi, Rym Oualha, Mohamed Aziz Cherif, Kobby Panford-Quainoo, Arnu PretoriusarXiv 2025 · DOI · arXiv

Introduction

Geospatial Foundation Models (GFMs) have promised a revolution in planetary monitoring, yet they often remain "ivory tower" artifacts—powerful in theory but prohibitively difficult to deploy in practical scenarios. A new framework from InstaDeep, InstaGeo, addresses this by providing an open-source, end-to-end pipeline that automates the transition from raw satellite data to interactive web deployment.

InstaGeo overcomes three primary bottlenecks in Earth Observation (EO): the Data Gap (complex preprocessing), the Complexity Gap (massive model sizes), and the Deployment Gap (lack of integrated visualization tools). By employing task-specific distillation, InstaGeo produces models up to 8x smaller than standard fine-tuned versions with minimal performance loss, enabling a complete "data-to-deployment" cycle in less than a single working day.

The InstaGeo Framework

The framework is structured into three modular components designed to unify previously siloed steps of the geospatial ML workflow.

Overview of the InstaGeo framework, illustrating the Data, Model, and Application components.
  1. Data Component (chip_creator): This module automates data curation by querying STAC-compliant catalogs (e.g., HLS, Sentinel-2), handling cloud masking, and generating model-ready GeoTIFF "chips."
  2. Model Component: Instead of full-scale fine-tuning, InstaGeo utilizes a teacher-student distillation approach to create "right-sized" models for specific tasks.
  3. Application Component: A built-in web interface (React/FastAPI) uses TiTiler to serve map tiles, allowing users to perform on-the-fly inference and visualize results interactively.

Task-Specific Distillation

A core innovation of InstaGeo is its approach to model efficiency. Standard fine-tuning often results in over-parameterized models for relatively simple EO tasks. InstaGeo addresses this through a teacher-student distillation mechanism.

The InstaGeo distillation process: a large teacher GFM guides a smaller student model using combined distillation and segmentation losses.

In this setup, a large, fine-tuned GFM (the teacher) guides a smaller student model composed of only the first $n$ layers of the teacher's encoder. The student is trained using a dual loss function: a distillation loss (KL Divergence) to match the teacher's probability distributions and a standard task loss (Cross-Entropy) against ground truth labels. This "right-sizing" ensures high performance with significantly reduced compute requirements.

Results and Impact

Validation across benchmarks such as flood mapping, crop segmentation, and locust breeding ground prediction demonstrates the framework's efficacy:

  • Efficiency: Distilled models achieved 2x to 8x size reductions. For flood mapping, compute requirements dropped from 49.3 to 27.5 GFLOPs with negligible accuracy loss.
  • Performance: InstaGeo achieved a new state-of-the-art mIoU of 60.65% on US crop segmentation, a 12 percentage point improvement over previous methods, largely due to superior automated data curation.
  • Operational Speed: The framework enables a single engineer to move from raw data to a functional web-map service in approximately 8.5 hours.

Conclusion

InstaGeo shifts the focus from massive model architectures to efficient, data-centric deployment. By automating the heavy lifting of data engineering and model compression, it lowers the barrier for practitioners to apply state-of-the-art geospatial AI to real-world humanitarian and environmental challenges.

Summary updated

Abstract

Open-access multispectral imagery from missions like Landsat 8-9 and Sentinel-2 has fueled the development of geospatial foundation models (GFMs) for humanitarian and environmental applications. Yet, their deployment remains limited by (i) the absence of automated geospatial data pipelines and (ii) the large size of fine-tuned models. Existing GFMs lack workflows for processing raw satellite imagery, and downstream adaptations often retain the full complexity of the original encoder. We present InstaGeo, an open-source, end-to-end framework that addresses these challenges by integrating: (1) automated data curation to transform raw imagery into model-ready datasets; (2) task-specific model distillation to derive compact, compute-efficient models; and (3) seamless deployment as interactive web-map applications. Using InstaGeo, we reproduced datasets from three published studies and trained models with marginal mIoU differences of -0.73 pp for flood mapping, -0.20 pp for crop segmentation, and +1.79 pp for desert locust prediction. The distilled models are up to 8x smaller than standard fine-tuned counterparts, reducing FLOPs and CO2 emissions with minimal accuracy loss. Leveraging InstaGeo's streamlined data pipeline, we also curated a larger crop segmentation dataset, achieving a state-of-the-art mIoU of 60.65%, a 12 pp improvement over prior baselines. Moreover, InstaGeo enables users to progress from raw data to model deployment within a single working day. By unifying data preparation, model compression, and deployment, InstaGeo transforms research-grade GFMs into practical, low-carbon tools for real-time, large-scale Earth observation. This approach shifts geospatial AI toward data quality and application-driven innovation. Source code, datasets, and model checkpoints are available at: https://github.com/instadeepai/InstaGeo-E2E-Geospatial-ML.git
BibTeX
@misc{yusuf2025instageo,
  author = {Ibrahim Salihu Yusuf and Iffanice Houndayi and Rym Oualha and Mohamed Aziz Cherif and Kobby Panford-Quainoo and Arnu Pretorius},
  title = {{InstaGeo: Compute-Efficient Geospatial Machine Learning from Data to Deployment}},
  year = {2025},
  eprint = {2510.05617},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.05617},
}
Figure