Introduction

Welcome to this solution, which brings together the full pipeline—from data ingestion and preprocessing to model training, quantization, and evaluation—into a clear, reproducible story.

Problem Overview

CubeSats provide cost‑effective astronomical sensing but face strict limits in size, weight, compute, and downlink bandwidth. The VERTECS mission, led by Kyushu Institute of Technology (Kyutech) and partners, studies the optical extragalactic background light (EBL). With a small‑aperture telescope and precise attitude control, VERTECS captures data that must be prioritized for transmission.

Limited onboard storage, processing capabilities, and slow communication can delay the return of vital data to Earth. Onboard machine learning helps the spacecraft pre‑select the most valuable images for downlink, maximizing the science return under constraints.

Objective

Develop a lightweight, efficient image classifier that accurately categorizes CubeSat imagery so high‑priority frames are identified for downlink under strict resource constraints. The approach balances computational efficiency with accuracy for resource‑limited platforms.

Five Quality Classes

• Blurry: Motion/defocus induced loss of detail.
• Corrupt: Sensor/transmission artifacts and stray light effects.
• Missing_Data: Partial frame loss or blank segments.
• Noisy: Radiation/electronic noise dominated frames.
• Priority: Clear scientifically valuable imagery to downlink first.

Pipeline Summary

The development workflow integrates global statistics normalization, model pruning, TFLite quantization, and evaluation in CPU‑constrained settings. This frontend adds interactive exploration and comparison to help contextualize predictions and quality metrics.

Frontend‑only changes — model and backend deployment remain unchanged.