Hyperparameter Tuning by Evolutionary Algorithm for Object Detection on Multimodal Satellite Imagery

Abstract

Deep learning workflow for object detection on multimodal satellite imagery from the modified VEDAI dataset and the related results are presented with a focus on the application of genetic algorithms (GA) for hyperparameter tuning for the deep learning (DL) models. The object detection is investigated for the three data input modality regimes, visible light (RGB) single modality, infrared (IR) single modality, and RGB+IR multimodal fusion (MF) modality, and the two methods of hyperparameter selections, “baseline” models with the pre-selected set of hyperparameters and “GA” models where hyperparameters tuned by GAs implemented in YOLOv5 DL model. The object detection performance of these DL models was measured by mean average precision (mAP). In general, for all models and modalities application of GA leads to a decrease of the mAP, and significant decay of generalization is observed for the single RGB and IR modalities in contrast to multimodal MF (RGB+IR) modality where generalization is preserved. These results demonstrate that the fusion of several modalities provides the class-specific synergistic data augmentation effect which provides better conditions for generalization than the standard and GA-tuned data augmentation methods.

Supported in part by the National Research Foundation of Ukraine (NRFU) grant 2022.01/0199.