An optimized training approach for meteor detection with an attention mechanism to improve robustness on limited data

Researchers have extensively used convolutional neural networks to detect meteor falls on Earth. However, when dealing with limited available data, these networks may need more robustness to classify new real-world images correctly. This study proposes an optimized training approach of a pre-trained model with an attention mechanism to achieve better generalization results in such a scenario. We compare two architectures, an optimized base model and another version with an attention mechanism. Furthermore, we present a new and publicly available optical meteor dataset that merges several public data sources. We used the merged dataset to train classification models combined with a stratified five-fold cross-validation strategy to determine the reliability of the prediction. The experimental results from both architectures showed good and similar performance. To further determine the best architecture, we performed an additional analysis with visual explanations in new observations. The architecture with an attention mechanism was the best model achieving a false alarm ratio of 2.6% and an accuracy of 97%.