Model Card

Overview

Lightweight CNN (pruned & quantized to TFLite) for 5-way quality classification in a CubeSat context: Blurry, Corrupt, Missing_Data, Noisy, Priority.

Intended Use

Onboard screening and prioritization of imagery before downlink. Not a substitute for final scientific quality vetting. Designed for low-latency CPU execution.

Problem Overview

CubeSats are effective but resource-constrained. Onboard inference helps prioritize high-value images under downlink and compute limits. This card consolidates the context so the home page can stay task-focused.

Data Summary

Training/validation/test splits derived from preprocessed NPY tensors (9,711 / 3,237 / 3,237). Class imbalance present (Priority most frequent; Corrupt & Missing_Data minority).

Preprocessing

• Global mean/variance statistics computed over training images.
• Resize to input resolution; bilinear interpolation.
• Normalization (/255 and statistical normalization as appropriate).

Model Efficiency

• Pruning schedule during training (progressive sparsity).
• Post-training quantization to TFLite.
• Mixed precision for faster convergence (training).

Limitations

• Class imbalance can reduce minority recall.
• Quantization may affect calibration.
• Distribution shift vs. real in-orbit conditions is possible.