Scaling LEO lightcurve classification: hyperparameter-tuned boosted decision trees on an expanded dataset

Building on previous low Earth orbit (LEO) satellite classification work, this paper aims to investigate the hyperparameter space of MMT-9 lightcurves and apply the optimal hyperparameters to classification. Both inter-platform and intra-platform lightcurve classification tasks are considered. The best hyperparameters were found to consist of 500 boosting steps, a learning rate of 0.01 and a max tree depth of 6 nodes. These hyperparameters were then used for classification with a lightcurve dataset size of 1000 lightcurves per platform, an order of magnitude greater than previous work, resulting in a slight increase in classification accuracy (87.92% versus 85.65%), suggesting a plateau was already being approached with 100 lightcurves per platform.