IceShadowNet

Deep Learning-Based Detection of Potential Subsurface Ice in Lunar South Polar Regions Using Chandrayaan-2 DFSAR Data: A Comparative Study of Custom CNN and ResNet (with Multimodal OHRC as Synthetic Future Work)

Motivated by ISRO's Chandrayaan-2 lunar ice characterization problem. This project extracts a focused, defensible Deep Learning research question from that domain rather than reproducing the full hackathon pipeline (GIS + radar processing + rover path planning).

---

1. Overview & Scope Clarification

IceShadowNet investigates whether deep learning models can distinguish anomalous polarimetric radar signatures from typical lunar regolith backscatter in lunar south-polar regions, using real Chandrayaan-2 DFSAR (Dual Frequency Synthetic Aperture Radar) L-band quad-polarimetric data.

Scope Decision (Real vs. Synthetic Data):

• Real-Data Pipeline: DFSAR-only. Real L1A complex slant-range (SLI) quad-polarimetric data (ch2_sar_ncxl_20260308) is processed to synthesize circular polarization channels and derive multilooked CPR and DOP labels.
• Multimodal (OHRC) Pipeline: Synthetic-only future work (synthetic demo, not part of the real-data benchmark). The downloaded real OHRC optical tile (ch2_ohr_ncp_20260719) does not geographically overlap the DFSAR acquisition strip (~35° longitude separation / ~1,800 km apart). Multimodal fusion architectures (EarlyFusionCNN, LateFusionResNet) are validated on synthetic data only and scoped as future work pending overlapping optical coverage.

2. Scientific Framing & Limitations

• Label Interpretation: Labels derived from $\text{CPR} > 1.0 \land \text{DOP} < 0.13$ define "regions with anomalous circular polarization and low degree of polarization" (consistent with multiple volume scattering / potential subsurface ice in lunar literature). Because no external PSR (Permanently Shadowed Region) or DEM slope mask is applied, labels denote anomalous CPR regions, NOT confirmed physical water-ice.
• Model Input Policy (Anti-Circularity): Model inputs are strictly raw radar log-intensities ($\log_{10}|S_{HH}|^2, \log_{10}|S_{HV}|^2, \log_{10}|S_{VV}|^2$). Derived quantities (CPR, DOP) are strictly excluded from network inputs to prevent the models from trivially memorizing the labeling rule.
• Multilook Averaging: Single-look complex radar data mathematically forces $\text{DOP} \equiv 1.0$ at every pixel. A $7 \times 7$ boxcar multilook filter is applied to the polarimetric covariance terms before calculating DOP and CPR.

3. Objectives

• Build a leak-free labeled patch dataset from real DFSAR L1A complex imagery using spatial block splitting.
• Implement a custom CNN from scratch and adapt ResNet18 for raw multi-channel SAR backscatter.
• Compare models on Accuracy, Precision, Recall, F1, ROC-AUC, training time, parameter count, and confusion matrix.
• Evaluate classical baselines (Plain Threshold and Decision Tree on CPR/DOP) on the identical spatial test split.
• Use Grad-CAM to evaluate whether models attend to physically plausible radar backscatter features.

4. Dataset

Source	Payload	Processing Level	Status / Use
Chandrayaan-2	DFSAR (L-band quad-pol)	L1A SLI (Complex float32)	REAL: Primary input; quad-pol complex scattering matrix ($S_{HH}, S_{HV}, S_{VH}, S_{VV}$)
Chandrayaan-2	OHRC (Panchromatic optical)	Calibrated L1	FUTURE WORK: Disjoint footprint; preserved for future multimodal alignment (synthetic demo, not part of the real-data benchmark)

• Real DFSAR tile: ch2_sar_ncxl_20260308t062017233 (South Polar strip, −81.7° to −86.3° latitude).
• Label rule: $\text{CPR} > 1.0 \land \text{DOP} < 0.13$ with multilook window $W=7$ ($\ge 5$ candidate pixels per $64 \times 64$ patch).
• Split strategy: Spatial block split grid $(1, 20)$ along the ground track to prevent adjacent patch leakage.

5. Methodology

Real DFSAR L1A Complex SLI (HH, HV, VH, VV)
        ↓
Multilook Polarimetric Analysis (W=7 boxcar)
   ├── S_RR = 0.5 * ((Shh - Svv) + 2i*Shv)   [Same-sense circular]
   └── S_RL = 0.5 * (Shh + Svv)              [Opposite-sense circular]
        ↓
Label Generation: CPR > 1.0 & DOP < 0.13 (min 5 px/patch)
        ↓
Model Input Construction: Raw Log-Intensities [HH, HV, VV]
(CPR and DOP strictly excluded from inputs)
        ↓
Spatial Block Split (Grid 1x20, zero spatial leakage)
        ↓
Model Training (CustomCNN, ResNet-scratch, ResNet-pretrained) & Baselines (Threshold, Decision Tree)
        ↓
Evaluation & Grad-CAM Explainability

6. Results

Real-Data Benchmark & Analysis:
For the completed real Chandrayaan-2 DFSAR benchmark results and scientific interpretation, see:

• Section 3 Analysis Draft: Headline finding (no deep model beats Logistic Regression on pooled metrics), CustomCNN vs ResNet_scratch parameter efficiency tie, ImageNet pretraining transfer breakdown, per-strip prevalence dynamics, and baseline sensitivity.
• Real-Data Benchmark Results (Spatial Block-Bootstrap CIs): Default and conditional 95% CIs across all 15 deep learning runs (seeds 42, 123, 2024) and classical baselines on full and clean test sets.
• Along-Track Per-Strip Metrics Breakdown: Granular strip-by-strip (strips 6, 10, 14) breakdown and zero-TP detection analysis.
• Architecture Comparison: Cross-variant summary including ResNet_pretrained v2a/v2b.
• Interactive Web Dashboard: Interactive visual exploration of benchmark metrics, pipeline stages, and patch classifications. Live dashboard: (add URL after deploy)

Historical Synthetic Pipeline Benchmark:
(synthetic demo, not part of the real-data benchmark)
Demonstrates initial pipeline execution across multiple seeds and epochs on synthetic data.

Model	Accuracy	Precision	Recall	F1	ROC-AUC	Support	Params	Time (s)
CustomCNN	0.7361±0.0196	0.4444±0.3143	0.1905±0.1347	0.2667±0.1886	0.6751±0.0555	TP:2 FP:1 FN:5 TN:16 (Pos:7 Neg:17)	618,209	76.0
ResNet (Scratch)	0.8194±0.0393	0.7361±0.0856	0.6190±0.1347	0.6605±0.0915	0.8459±0.0914	TP:3 FP:1 FN:4 TN:16 (Pos:7 Neg:17)	11,173,889	151.1
ResNet (Pretrained)	0.7778±0.0196	0.6488±0.0748	0.5714±0.1166	0.5945±0.0521	0.8403±0.0562	TP:4 FP:3 FN:3 TN:14 (Pos:7 Neg:17)	6,067,201	73.9
EarlyFusionCNN (synthetic demo)	0.7222±0.0196	0.2222±0.3143	0.0952±0.1347	0.1333±0.1886	0.4594±0.0309	TP:0 FP:0 FN:7 TN:17 (Pos:7 Neg:17)	618,497	54.0
LateFusionResNet (synthetic demo)	1.0000±0.0000	1.0000±0.0000	1.0000±0.0000	1.0000±0.0000	1.0000±0.0000	TP:7 FP:0 FN:0 TN:17 (Pos:7 Neg:17)	22,606,273	200.1

7. Experiments

Experiment	Input	Status
CNN	DFSAR only	Real-data benchmark complete
ResNet	DFSAR only	Real-data benchmark complete
CNN (Early Fusion)	DFSAR + OHRC (multimodal)	Synthetic demo, not part of the real-data benchmark
ResNet (Late Fusion)	DFSAR + OHRC (multimodal)	Synthetic demo, not part of the real-data benchmark

8. Tech Stack

• Language: Python
• DL Framework: PyTorch
• Geo/Radar processing: GDAL, rasterio, NumPy, SciPy
• Visualization: Matplotlib, QGIS (for geospatial output)
• Explainability: Grad-CAM

9. Repository Structure

IceShadowNet/
├── data/
│   ├── raw/
│   │   └── MISSING_DATA.md      # PRADAN download guide for real DFSAR L1A tiles
│   └── processed_real/          # Spatial NPZ splits (train, val, test)
├── models_real/                 # PyTorch model weights for 15 real benchmark runs
├── reports/
│   ├── real/                    # Benchmark metrics, comparison tables, and analysis
│   │   ├── section3_analysis.md # Section 3 draft: findings, comparisons, limitations
│   │   ├── results_real_ci_all.md # 3-block bootstrap CIs (default & conditional)
│   │   ├── per_strip_metrics.md # Granular along-track per-strip metrics (strips 6, 10, 14)
│   │   ├── results_real_variants_comparison.md # Side-by-side comparison across all variants
│   │   ├── trivial_baselines_and_ambiguity_note.md # All-positive baselines & ambiguity note
│   │   └── training_summary*.json # Stored metrics for programmatic reproduction
│   └── synthetic_archive/       # Archived placeholder synthetic reports
├── scripts/
│   ├── train_all.py             # Main DFSAR multi-seed training pipeline
│   ├── baseline_real.py         # Decision Tree & Logistic Regression feature baselines
│   ├── ingest_and_split.py      # Real DFSAR L1A tile ingestion & spatial block splitting
│   ├── run_gradcam.py           # Grad-CAM heatmap visualization runner
│   ├── _smoke_gpu.py            # 1-epoch GPU verification smoke test
│   ├── train_multimodal.py      # Multimodal training (synthetic demo, not part of real benchmark)
│   ├── ingest_multimodal.py     # Multimodal ingestion (synthetic demo, not part of real benchmark)
│   ├── verify_real_tile.py      # DFSAR L1A GeoTIFF inspector
│   └── analysis/                # Standalone analysis & metric reproduction tools
│       ├── generate_comparison.py
│       ├── compute_per_strip.py
│       ├── generate_phase4_reports.py
│       ├── compute_trivial.py
│       ├── generate_ci_fixed.py
│       └── audit_bootstrap.py
├── src/
│   ├── datasets/                # IcePatchDataset & spatial block split logic
│   ├── models/                  # CustomCNN, IceResNet, MultimodalFusion
│   ├── preprocessing/           # Multilooking, circular polarization, CPR/DOP
│   ├── train.py                 # Core training loop & optimizer routines
│   ├── evaluate.py              # Spatial block-bootstrap CI & metric evaluation
│   └── gradcam.py               # Grad-CAM implementation
├── tests/
│   └── test_pipeline.py         # Pytest test suite (splitting, circularity, data ops)
├── .github/workflows/tests.yml  # GitHub Actions CI workflow
├── README.md
├── PRD.md
├── requirements.txt
└── config.yaml

10. Setup

git clone https://github.com/nish-debug15/IceShadowNet
cd IceShadowNet
pip install -r requirements.txt

Data access: register at PRADAN, download DFSAR/OHRC tiles for the target crater, place under data/raw/.

11. Team

Role	Responsibility
Data & Labeling	DFSAR/OHRC acquisition, CPR/DOP labeling, spatial split
Custom CNN	Architecture design, training, tuning
ResNet	Adaptation for SAR input, transfer-learning decisions, training
Evaluation & Explainability	Metrics, confusion matrix, Grad-CAM
Integration & Report	Pipeline glue, geospatial mapping, report/demo assembly

12. Acknowledgement

This research is based partially on the results obtained from the Chandrayaan-2, second lunar mission of the Indian Space Research Organisation (ISRO), archived at the Indian Space Science Data Centre (ISSDC).

13. References

• ISRO/ISSDC Chandrayaan-2 mission data — https://pradan.issdc.gov.in/ch2/
• ISRO Lunar Polar Region overview — https://www.isro.gov.in/Lunar_Polar_Region.html
• He et al., Deep Residual Learning for Image Recognition (ResNet), 2015
• Selvaraju et al., Grad-CAM: Visual Explanations from Deep Networks, 2017