Geospatial Hazard Detection

TerraScan

Upload multispectral terrain imagery or coordinate targets
to run the U-Net + 1D-CNN predictive pipeline.

0.81 Dice Score

91.3% CNN Accuracy

0.964 ROC-AUC

Initialize Scan Protocol

Drop satellite imagery (RGB / TIFF) · click to browse

Detection Threshold 0.50

Overlay Opacity 55%

Pipeline Execution

⬇️Ingest

🧠U-Net

📐Extract

⚡1D-CNN

🗺️Fusion

Raw S-2 Input

U-Net Segmentation

CNN Risk Heatmap

Synthesized Overlay

Computed Hazard Probability —

—

Awaiting analysis output

Affected Area

—

Confidence μ

—

Sigmoid Thr.

—

Kothamangalam

10.0600°N · 76.6220°E

Munnar Sector

10.0889°N · 77.0595°E

Bijie Primary

27.4700°N · 106.5700°E

Latitude (N/S)

Longitude (E/W)

System Architecture

U-Net Segmentation (ResNet-34)

Encoder pretrained on ImageNet. Trained with Dice + BCE loss to handle severe class imbalance (~8% landslide pixels). Outputs a binary landslide mask isolating immediate hazard zones.

Conditioning Factor Extraction

Derives six critical geospatial factors per pixel: Elevation, Slope, and Aspect from Copernicus DEM, integrated with Texture, pseudo-NDVI, and Brightness from multispectral RGB channels.

1D-CNN Risk Scoring

Lightweight 1D convolutional network trained on balanced tabular data. Processes extracted factors to output a continuous risk probability scalar [0, 1] per spatial pixel.

Data Fusion & Visualization

U-Net binary masks and CNN risk tensors are blended into a GIS-ready topographic heatmap, optimizing field team dispatch and early warning protocols.