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

TL;DR

In geospatial ML, the adoption of large, pre-trained foundation models is hindered by the lack of open-source data pipelines and deployment tooling. InstaGeo provides an end-to-end framework that automates data curation, enables task-specific distillation to produce compact models, and deploys results via an interactive web-map, all using open-source components. The approach reproduces and extends published benchmarks, achieving near-baseline performance with dramatically reduced model size (up to eightfold fewer parameters) and faster data-to-deployment cycles, including a state-of-the-art crop segmentation result on an expanded CDL dataset. By delivering a low-carbon, real-time solution from labeled observations to actionable maps, InstaGeo lowers the barrier to operational geospatial AI and promotes data-driven, application-focused geospatial research.

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

Figures (5)

• Figure 1: Overview of InstaGeo for geospatial machine learning. (a) Data component. Point or raster observations are ingested, matched to satellite granules via STAC API search, grouped by common acquisition, cut into multitemporal “chips”, with labels rasterised into segmentation maps, cloud‑masked and serialised as GeoTIFF pairs. (b) Model component. A frozen GFM encoder plus segmentation head acts as the teacher; a shallower student shares the first $n$ encoder layers. The student is trained with a combined segmentation loss (ground‑truth maps) and distillation loss (KL divergence to the teacher logits), yielding a lightweight task‑specific model. (c) Application component. The trained model runs inference on new sampled images to generate prediction rasters, which are then served via a tile server using TiTiler and rendered as interactive web‑map layers. Together, these three components convert raw observations into operational geospatial intelligence in a single, reproducible workflow.
• Figure 2: InstaGeo reproduces unpublished data pipelines for published datasets. For each task we the following performances: (i) the Baseline reported in the original study, (ii) the InstaGeo-Baseline trained using the authors’ original dataset but using InstaGeo's model component, and (iii) the InstaGeo-Replica trained on a dataset entirely reconstructed using InstaGeo. (a) Performance comparison for the flood mapping task. The original dataset was created using Sentinel-2 imagery, and we created two replicas, one from HLS and another from Sentinel-2. b Performance comparison for the multi-temporal crop segmentation task. (c) Performance comparison for the desert locust breeding ground prediction task.
• Figure 3: Task-specific distillation yields lightweight models with comparable performance and reduced computational cost. Each column corresponds to a downstream task, where the top row compares the performance of a vanilla fine-tuned model (Vanilla FT) against a lightweight student model obtained via task-specific distillation across four evaluation metrics: mIoU, mAcc, mF1, and ROC-AUC. The bottom row illustrates the model characteristics by comparing the parameter counts (in millions) and estimated compute requirements (in GFLOPs) of the two models. Despite significantly reduced model size and compute, task-specific student models retain comparable performance, demonstrating the effectiveness of task-specific distillation.
• Figure 4: Impact of larger CDL variants on multitemporal crop segmentation, showing mIoU, mAcc, mF1 and ROC–AUC for the published baseline (3.8k chips), expanded 2022 CDL (14k chips) and 2024 CDL (18k chips)
• Figure 5: Lightweight web-map application - Application component. (a) The user delineates a region of interest by drawing/editing bounding boxes and then selects the model, date, and other input parameters. (b) Submitting the request spawns data‑processing, inference and visualization preparation jobs that are queued for execution. (c) Job progress is tracked in the Tasks Monitor panel. (d) Upon completion, predictions are rendered interactively on the map.