Skip to main content

MS Cloud Detection Benchmark

Bandwidth Efficiency

Odin v0 processes a full 120-megapixel Sentinel-2 scene in 29.88 s at 10.1 W — 2.1× faster than an equal-configuration Jetson baseline and 29.9% lower system power. Deterministic tile inference (σ = 0.31 ms) enables predictable onboard data pipelines.

Executive Summary​

MetricOdin v0 (INT8)NVIDIA Jetson (TRT FP16)Delta
System Power10.1 W14.4 W29.9% lower
Latency Jitter (σ)0.31 ms0.64 ms2.1× more deterministic
Scene Processing Time29.88 s60.34 s2.1× faster
Inference Rate94 it/s42 it/s2.2× higher
Accuracy (F2-Score)0.790.81Negligible loss

The Bandwidth Bottleneck​

Downlinking data from LEO is expensive and capacity-constrained. Up to 55% of satellite imagery is unusable due to cloud cover, wasting valuable downlink bandwidth.

Odin v0 enables On-Orbit Cloud Filtering:

  1. Detect: Pixel-level cloud and shadow segmentation at 10m resolution.
  2. Filter: Cloud-contaminated tiles are flagged before downlink.
  3. Reduce: Only clear-sky tiles are prioritized, eliminating wasted downlink capacity.

Pipeline Architecture​

MS Cloud Detection — U-Net Pipeline ArchitectureS-2InputNorm÷10KTile224²EncoderMBConv16·112²MBConv24·56²MBConv32·28²BtInck320·7²skipDecoder + OutputUp+Cat16·112²Up+Cat32·28²Up+Cat96·14²Binary0 / 1ORFusionSceneMask2,809tilesFull U-Net inference · normalisation and mask fusion on host

Implementation​

The complete DTACSNet-CD model (U-Net with MobileNetV2 encoder) runs on the D-IMC accelerator. Host CPU handles normalization and mask fusion only.

StageDescription
NormalizationDivide by 10,000 (reflectance normalization)
Full U-Net InferenceEncoder-decoder, INT8, all 2,809 tiles
Mask FusionBoolean OR-fusion into full-scene cloud mask

Tiling Strategy​

Full Sentinel-2 scenes (10,980 × 10,980 px) processed with sliding-window strategy:

  • Tile size: 224 × 224
  • Total tiles per scene: 2,809
  • Inference rate: 94 tiles/s (Odin v0) vs 42 tiles/s (Jetson batch=1)

Conclusion: Benefits for Space DPUs​

  • Power Budget: At 10.1 W, the pipeline fits within a standard 15 W payload allocation — leaving margin for radio and housekeeping. Jetson-only at 14.4 W exceeds a 10 W inference budget.
  • Deterministic Streaming: σ = 0.31 ms per tile ensures consistent throughput for downstream atmospheric correction and change-detection pipelines.
  • Radiation Surface: Full model runs on D-IMC accelerator — GPU utilization is minimal during inference, reducing the SEU-sensitive compute surface.