README.md

metadata

language:
  - en
license: cc0-1.0
tags:
  - tabular-regression
  - spatial-ml
  - radiation
  - safecast
  - environmental-science
  - anomaly-detection
  - pytorch
  - safetensors
  - onnx
datasets:
  - hsilvosa/safecast-radiation
metrics:
  - r2
  - rmse
  - mae
model-index:
  - name: safecast-spatial-harmonic-net
    results:
      - task:
          type: tabular-regression
          name: Global Radiation Spatial Regression
        dataset:
          type: hsilvosa/safecast-radiation
          name: Safecast Historical Radiation Measurements
        metrics:
          - type: r2
            value: 0.0785
            name: R2 Score (log-scale)
          - type: rmse
            value: 3.11636
            name: RMSE (uSv/h)
          - type: mae
            value: 0.67873
            name: MAE (uSv/h)

Global Radiation Anomaly Map & Spatial Regressor (Safecast)

Model Overview

This repository provides SpatialHarmonicNet, a continuous spatial neural regression model that predicts environmental ambient radiation levels in microsieverts per hour (uSv/h) anywhere on Earth from geographic coordinates (latitude, longitude) and performs real-time radioactive anomaly detection.

The model is trained on crowdsourced radiation sensor measurements from the Safecast Historical Dataset, which spans over 265 million measurements collected worldwide from 2011 to 2026.

Architecture

SpatialHarmonicNet is designed specifically for spherical planetary coordinates:

1. Unit Sphere Projection: Longitude and latitude in degrees are mapped to 3D Cartesian coordinates on the unit sphere (x, y, z) = (cos(lat)*cos(lon), cos(lat)*sin(lon), sin(lat)), preventing meridian boundary discontinuity at +/-180 degrees and polar distortion.
2. Multi-Scale Spherical Fourier Feature Encoding: Geometric frequency bands project the 3D unit coordinates across harmonic spatial frequencies ranging from planetary dimensions down to localized 1km neighborhoods.
3. Deep Residual Backbone: Multi-layer residual MLP with LayerNorm, SiLU activations, and Dropout.
4. Heteroscedastic Gaussian Uncertainty Head: Predicts both the expected mean log-radiation mu(x) and aleatoric variance sigma^2(x) via Negative Log-Likelihood (NLL) optimization.
5. Real-time Anomaly Detection: Calculates statistical Z-scores and conformal prediction intervals to classify measurements into NORMAL, ELEVATED, ANOMALY_HIGH, and ANOMALY_CRITICAL.

Evaluation Results

Evaluation performed on a holdout spatial test split (stratified across global 0.1-degree spatial grid cells):

Metric	SpatialHarmonicNet (PyTorch)
R2 Score (log-scale)	0.0785
RMSE (uSv/h)	3.11636 uSv/h
MAE (uSv/h)	0.67873 uSv/h
95% Confidence Interval Coverage (PICP)	94.3%

Reference Landmark Verification

Location	Category	Expected / Measured Baseline	Anomaly Trigger (at 5.0 uSv/h)
Chernobyl Reactor 4 Shelter	Nuclear Exclusion Zone	Elevated	ANOMALY_HIGH / CRITICAL
Pripyat Red Forest	Nuclear Exclusion Zone	Elevated	ANOMALY_HIGH / CRITICAL
Fukushima Daiichi	Nuclear Exclusion Zone	Elevated	ANOMALY_HIGH / CRITICAL
Tokyo Metropolitan Area	Urban Background	~0.05 - 0.08 uSv/h	ANOMALY_CRITICAL (Z > 12)
Paris, France	Urban Background	~0.06 - 0.09 uSv/h	ANOMALY_CRITICAL (Z > 12)
New York City, USA	Urban Background	~0.07 - 0.10 uSv/h	ANOMALY_CRITICAL (Z > 12)
Denver, USA (Mile-High)	Elevated Cosmic Background	~0.12 - 0.16 uSv/h	ANOMALY_CRITICAL (Z > 10)

Quickstart & Usage

1. Installation

pip install torch safetensors numpy pandas scipy

2. Python Inference Example

import json
import torch
from safetensors.torch import load_file
from radiation_map.models.spatial_net import SpatialHarmonicNet
from radiation_map.models.anomaly_detector import RadiationAnomalyDetector

# 1. Initialize model
model = SpatialHarmonicNet(num_frequencies=32, max_frequency_log=4.5, hidden_dims=(256, 256, 128, 64))
state_dict = load_file("model.safetensors")
model.load_state_dict(state_dict)
model.eval()

# 2. Predict baseline radiation at a coordinate
# Coordinates for Tokyo (35.6895 N, 139.6917 E)
pred = model.predict_radiation(latitudes=35.6895, longitudes=139.6917)
print(f"Predicted baseline: {pred['radiation_usv']:.4f} uSv/h")
print(f"95% Confidence Interval: [{pred['ci_lower_usv']:.4f}, {pred['ci_upper_usv']:.4f}] uSv/h")

# 3. Real-time Anomaly Detection
detector = RadiationAnomalyDetector(model)
result = detector.detect(
    latitude=35.6895,
    longitude=139.6917,
    observed_value=2.50, # hypothetical spike in uSv/h
    unit="usv"
)

print(f"Severity: {result.severity.value}")
print(f"Z-score: {result.z_score:.2f}")
print(f"Fold increase: {result.fold_increase:.1f}x")
print(f"Description: {result.description}")

3. ONNX Runtime Inference

import numpy as np
import onnxruntime as ort

session = ort.InferenceSession("spatial_regressor.onnx")

# Project lat/lon to 3D Cartesian coordinates
lat, lon = np.radians(35.6895), np.radians(139.6917)
xyz = np.array([[np.cos(lat)*np.cos(lon), np.cos(lat)*np.sin(lon), np.sin(lat)]], dtype=np.float32)

inputs = {"coords_cartesian": xyz}
mu_log, log_var = session.run(None, inputs)

# Inverse log transform to get uSv/h
pred_usv = np.expm1(mu_log[0][0]) / 10.0
print(f"ONNX Predicted uSv/h: {pred_usv:.4f}")

Intended Use & Limitations

• Intended Use: Environmental baseline modeling, spatial regression research, citizen-science data exploration, and screening for radioactive anomalies.
• Limitations: Safecast data is crowdsourced and collected with mobile bGeigie Geiger-Muller counters. It is not official regulatory or government monitoring data. Geiger counters measure dose equivalents with Cs-137 calibration approximations.

Citation & Attribution

@misc{safecast_spatial_radiation,
  author = {Safecast Contributors and Project Authors},
  title = {Global Radiation Anomaly Map and Spatial Regressor},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/models}}
}