Evaluation of Parameter-Scaling for Efficient Deep Learning on Small Satellites