STARC: comparação de CNNs com variações de entrada para classificação de trilhas de satélite em imagens astronômicas

Abstract

This work compared two classification pipelines for automatic detection of satellite trails in wide-field astronomical images from the MeerLICHT survey. A set of 356 images (178 images and 178 masks) was used, resized to 2,048×2,048 and sliced into 224×224 patches (9×9 grid). The raw, Hough, and combined input modalities were evaluated using two pre-trained convolutional architectures (ResNet-18 and DenseNet-121). Patch labeling was derived from the masks; the data was grouped by image into 70/15/15 for training/validation/testing; Focal Loss was adopted as the loss function; and the data augmentation policy promoted discrete rotations (90°/180°/270°) only in the positive patches during training. Model selection considered balanced accuracy, with threshold scanning in validation and application of the optimal threshold in testing. The results were generally positive: the combined modality presented the best compromise between sensitivity and specificity (up to 91.4% balanced accuracy with ResNet-18), the raw modality maintained solid and stable performance, and the isolated Hough modality fell short due to false positives. The differences between ResNet-18 and DenseNet-121 were modest, while the computational cost of DenseNet-121 was higher, consolidating ResNet-18 in combined as the recommended configuration for practical use. Although focused on satellite tracks, the pipeline showed potential for being extended to tasks involving the classification of astronomical artifacts in general.