nafizrahaman/geopulse-gpft-mini

GeoPulse GPFT-mini checkpoints

Weights for GeoPulse, an open-source multimodal Earth-change mapper (flood, wildfire/burn, vegetation disturbance) built on Sentinel-1, Sentinel-2 and the Copernicus DEM.

pip install https://github.com/nrahaman1/geopulse/archive/refs/heads/main.zip
geopulse models pull                      # downloads these checkpoints and verifies their SHA-256
geopulse infer --task wildfire --aoi aoi.geojson --before 2024-12-01/2025-01-06 --after 2025-02-01/2025-03-31

Each <model>.pt has a <model>.json card (tasks, training datasets and manifest hashes, label sources, validation metrics, temperature, checkpoint SHA-256). Test-set reports are in eval/.

Model card: GeoPulse v0.2 (flood, wildfire, vegetation disturbance)

Checkpoints in models/, each with a JSON card beside it (tasks, checkpoint SHA-256, dataset manifest SHA-256, label sources, git SHA, validation metrics, temperature). Full training records: runs/<model>/run.json. Evaluation reports: runs/<model>/eval_<task>_test.{json,md}.

Model	Tasks	Trained on	Sensor dropout
threshold-baseline	all	nothing (physics)	—
gpft-flood-mini / -nodrop	flood	GeoPulse-Bench flood	yes / no
gpft-wildfire-mini / -nodrop	wildfire	GeoPulse-Bench wildfire	yes / no
gpft-vegetation-mini / -nodrop	vegetation	GeoPulse-Bench vegetation	yes / no
gpft-multitask-mini	all three	all three benchmarks	yes

auto resolves to the newest checkpoint that supports the requested task (currently gpft-multitask-mini).

Intended use

Research and demonstration of multimodal change mapping from Sentinel-1 + Sentinel-2 + Copernicus DEM, run locally on AOIs of up to a few hundred km². Outputs are probabilities with an uncertainty map and an explicit "review recommended" area; they are not an operational emergency or forest-monitoring product.

Out of scope: operational use without expert review; flooded vegetation or dense urban flooding (flood labels cover open water only); non-stand-replacing forest degradation, crop anomalies (vegetation labels are Hansen stand-replacing loss); burn severity as a learned output (severity classes are dNBR thresholds, not model output); resolutions other than 10 m.

Architecture

Per-sensor adapters shared across time → per-pixel quality-gated fusion (unobserved or cloudy sensors get exactly zero weight) → change encoder [pre, post, post−pre, |post−pre|] + terrain → 3-level U-Net (GroupNorm) → one logit per task + a shared change logit, ÷ temperature. Uncertainty = entropy of the mean of 8 MC-dropout passes. ~0.54 M parameters, trained from scratch. Post-processing constraint for burn and disturbance: pixels that were open water before the event are set to 0 (a model never shown the sea once mapped ocean as burned).

Training data

Benchmark	Events (train / val / test)	Tiles	Labels
flood	Missouri 2019, Harvey 2017, Idai 2019, Florence 2018, Ian 2022 / Emilia-Romagna 2023 / Pakistan 2022	236	weak: S1+S2 consensus
wildfire	Carr, Camp, Ranch 2018; Thomas, Tubbs 2017 (California) / Eagle Creek 2017 (Oregon) / Martin 2018 (Nevada sagebrush)	212	MTBS burn severity (low–high = burned)
vegetation	Georgia pines 2019, Oregon Coast 2020, Hurricane Michael 2018 / Västerbotten, Sweden 2019 / Pará, Brazil 2020	176	Hansen GFC loss year

Each test event is out of distribution: a continent (flood), an ecoregion (wildfire) or a biome (vegetation) absent from training.

Label caveats. Flood labels are weak: scores measure agreement with the two-sensor consensus, not ground truth, and the flood baseline cannot be scored against them (circular). MTBS and Hansen are independent of the Sentinel inputs (Landsat-derived, analyst-thresholded), so baseline-vs-model comparisons on wildfire and vegetation are fair. MTBS "unburned to low" and Hansen loss in other years are ignored, not treated as negatives.

Results — held-out test events, one seed

IoU against the benchmark labels; sensors removed at inference by zeroing their validity masks.

Flood — Pakistan 2022 (weak labels)

Model	S1+S2	S1 only	S2 only	ECE
gpft-flood-mini	0.940	0.922	0.887	0.037
gpft-flood-mini-nodrop	0.975	0.957	0.911	0.016
gpft-multitask-mini	0.959	0.685	0.969	0.030

Wildfire — Martin Fire 2018, Great Basin sagebrush (MTBS)

Model	S1+S2	S1 only	S2 only	ECE
threshold-baseline (dNBR)	0.901*	0.013	0.901*	0.070
gpft-wildfire-mini	0.017	0.371	0.016	0.315
gpft-wildfire-mini-nodrop	0.232	0.445	0.323	0.267
gpft-multitask-mini	0.707	0.338	0.765	0.107

* The baseline first scored 0.076: its NDVI > 0.2 "fuel" gate excluded sagebrush steppe (pre-fire NDVI ≈ 0.17). The gate was lowered to 0.10 after diagnosing this on the test event, so 0.901 is not a clean held-out number. On the validation fire (Eagle Creek, conifer) the baseline scores 0.840 and the learned models 0.947–0.955.

Vegetation disturbance — Novo Progresso, Pará 2020 (Hansen GFC)

Model	S1+S2	S1 only	S2 only	ECE
threshold-baseline (NDVI + VH drop)	0.212	0.113	0.188	0.060
gpft-vegetation-mini	0.496	0.103	0.525	0.025
gpft-vegetation-mini-nodrop	0.521	0.081	0.566	0.026
gpft-multitask-mini	0.560	0.200	0.557	0.021

Findings

1. Multi-task training helped every task's out-of-distribution test (design RQ7), measured against the matching single-task model with sensor dropout: flood 0.940 → 0.959, wildfire 0.017 → 0.707, vegetation 0.496 → 0.560. The shared encoder sees far more varied land cover (arid Sindh, Mozambique, Amazon), which plausibly explains the wildfire jump. Cost: the multi-task model's SAR-only flood score fell to 0.685.
2. Physics beat learning across an ecoregion shift for burns. Trained on five California forest/chaparral fires, the single-task models learned "burned forest" and missed low-biomass rangeland burns; dNBR transfers. In-distribution (conifer val fire) the learned models win by ~0.11 IoU.
3. Learning beat physics for vegetation disturbance by ~2.5×, where fixed NDVI/VH thresholds cannot separate stand-replacing loss from drought browning, partial windthrow and regrowth.
4. Sensor dropout (RQ3) did not help in any task: the no-dropout ablation matched or beat it with both sensors and in most single-sensor rows. Plausible reasons: few training events, and labels concentrated on pixels where both sensors carry the signal.
5. SAR-only burn mapping is weak for every model (≤ 0.45 IoU): C-band sees fire only through canopy loss.

Known limitations

• One seed and one test event per task: no confidence intervals; differences of a few IoU points are not evidence.
• MTBS covers CONUS through 2018 only; Hansen labels only stand-replacing loss, annually.
• Change labels are conservative S1+S2 consensus; the change head is secondary.
• Sentinel-2 SCL can mask dark floodwater as cloud shadow; winter scenes can leave large optical gaps (Camp Fire: 55 % valid), which SAR fills with explicitly higher uncertainty.
• Vegetation requests must compare the same season in different years; others get a warning.

Systems (RTX 4060 Laptop GPU, torch 2.11 + CUDA 12.8, batch 16, 256 px tiles)

FP32 10.1 ms/tile, 1.9 GB peak; FP16 autocast is not faster for this small model. Training: ~12 min per single-task model (40 epochs), ~25 min for the multi-task model (25 epochs). End-to-end inference for 100–200 km² is dominated by data access (≈1.5 min cold, ≈10–15 s cached).

License

Code and weights: Apache-2.0. Data and labels: see DATA_LICENSES.md.