NOAH TENET AI
23/02/2026
Noah's Ark
Quantum Tech Lab
Quantum Tech Lab
NOAH TENET AI
ITU Ingenuity Cup 2026 — Track 3 Submission
Urban Intelligence From Space
NOAH TENET AI leverages deep learning satellite imagery analysis to detect urban heat islands and built-up areas — contributing to resilient, sustainable cities worldwide.
Architecture
U-Net Deep Learning Model
// NETWORK ARCHITECTURE
ENCODER — Conv2D (3→64) + BN + ReLU
ENCODER — Conv2D (64→128) + MaxPool
ENCODER — Conv2D (128→256) + MaxPool
ENCODER — Conv2D (256→512) + MaxPool
↓
BOTTLENECK — Conv2D (512→1024)
↑
DECODER — ConvTranspose + Skip Connection
DECODER — ConvTranspose + Skip Connection
DECODER — ConvTranspose + Skip Connection
DECODER — ConvTranspose + Skip Connection
↓
OUTPUT — Conv2D (64→2) — Binary Segmentation
Semantic Segmentation
The U-Net architecture was chosen for its exceptional performance on satellite image segmentation tasks. Its encoder-decoder structure with skip connections preserves fine spatial details critical for pixel-level urban area detection.
- Gaofen-2 satellite imagery (512×512 px)
- Binary segmentation: built-up vs non-built-up
- Red channel detection (RGB 255, 0, 0 masks)
- CrossEntropyLoss + Adam optimizer (lr = 1e-4)
- 10 epochs training on dual GPU T4 ×2
- IoU + Overall Accuracy evaluation metrics
Visualization
Segmentation Demo
INPUT — GAOFEN-2 SATELLITE
OUTPUT — BUILT-UP MASK
Performance
Model Results
IoU
Intersection over Union
Primary segmentation metric measuring pixel-level overlap between predicted and ground truth built-up area masks.
T4 ×2
GPU Acceleration
Dual NVIDIA Tesla T4 GPUs enabling high-throughput parallel processing of large-scale satellite imagery.
GID
Gaofen Image Dataset
High-resolution Chinese Gaofen-2 satellite imagery covering diverse global urban and rural landscapes.
Model Training Completion
10 Epochs ✅
Training Dataset Coverage
27,300 Images
Test Predictions Generated
216 Masks ✅
ITU Submission Status
Succeeded ✅
Journey
Project Timeline
February 2026 — Phase I
Dataset Acquisition
Downloaded and processed the 12 GB GID (Gaofen Image Dataset) from the ITU challenge platform. Resolved dataset integrity issues and uploaded successfully to Kaggle cloud infrastructure for training.
February 2026 — Phase II
Model Development
Designed and implemented a custom U-Net architecture from scratch in PyTorch — no external segmentation libraries. Configured encoder-decoder with skip connections, batch normalization, and optimized data loaders for satellite imagery.
February 23, 2026 — Phase III
Training & Official Submission
Successfully trained the model on dual GPU T4 ×2 infrastructure via Kaggle Notebooks. Generated 216 binary segmentation masks for the full test set and submitted to the ITU Ingenuity Cup 2026 official platform.
✅ 2026-02-23 14:23:20 — Submission Succeeded
March 2026 — Phase IV
Round 2 — Thermal Analysis
If qualified: infrared and thermal band analysis for surface temperature mapping and urban heat island detection using advanced multispectral Gaofen satellite data processing algorithms.