ClearNet: Removing Clouds from Satellite Imagery with Diffusion

During the Indian monsoon, cloud cover renders up to 60% of LISS-4 satellite acquisitions unusable — right when agricultural surveillance matters most. ClearNet is the system our team (CloudXRecon) built for the Bharatiya Antariksh Hackathon 2026 to reconstruct those obscured regions.

The Problem

LISS-4, aboard ISRO's Resourcesat-2, is the highest-resolution optical sensor in the Indian remote sensing fleet at 5.8 m ground sampling distance. It feeds crop area estimation under PMFBY, flood mapping, and land records. But as an optical sensor, it is blind whenever clouds roll in — and the monsoon persists for 15-20 consecutive days, well beyond the 5-day revisit cycle.

Existing cloud removal methods target Sentinel-2's 10 m, 13-band format. None work natively with LISS-4's 3-band, 5.8 m GeoTIFF, and none enforce NDVI fidelity — the metric that actually matters for crop monitoring.

The Approach

We built a SAR-conditioned latent diffusion bridge. The core idea:

1. Cloud mask via Otsu-thresholded NSCI, splitting thin and thick cloud.
2. SAR fusion — co-register Sentinel-1 radar (cloud-penetrating) to LISS-4 geometry, fuse via multi-head cross-attention.
3. Diffusion bridge — instead of starting from Gaussian noise, the model transports directly from the cloudy distribution to the clear distribution. Cloud-free pixels are preserved exactly; only masked regions are reconstructed.
4. NDVI preservation loss — a spectral term that keeps the reconstructed NIR/Red ratio consistent with ground truth, so the output stays useful for agriculture.

Why a Bridge, Not a GAN

Conditional GANs collapse on thick-cloud scenes (above 40% coverage). The discriminator cannot tell hallucinated NIR reflectance from real reflectance, so it produces spectrally invalid fills. The diffusion bridge formulation sidesteps this by modelling the stochastic transport between distributions directly — no adversarial game, no mode collapse.

Domain Adaptation

The base model pre-trains on SEN12MS-CR (180,662 Sentinel-1/2 scene pairs). To adapt to LISS-4, we inject LoRA rank-8 adapters into the cross-attention layers only — cutting trainable parameters from ~85M to 1.2M. This lets us fine-tune on 200-500 LISS-4 pairs from Bhoonidhi without overfitting, inside the hackathon's compute window on a single A100.

Results

Under the 10-minute demo baseline (64 channels, 1.2M params, 10 epochs):

• PSNR: 19.59 dB (target: >31 dB at full scale)
• SSIM: 0.584 (target: >0.910)
• NDVI correlation: 0.402 (target: >0.95)

The full-scale configuration is projected to hit >31 dB PSNR — competitive with the published state of the art on SEN12MS-CR, while being the first method natively designed for LISS-4 and Bhuvan workflows.

What I Worked On

My role was systems and deployment — pipeline orchestration, inference optimisation, the demo interface, and submission packaging. The ML architecture was led by Kavish Tiwari, remote sensing and data by Ritesh Singh, and evaluation by Harshit Singh.

The full proposal with the architecture diagrams, equations, and quantitative comparison tables is at vxky.me/bah2026-proposal.