AERONET: Design and Evaluation of Systematic Experimental Frameworks for Custom CNN-Based Military Aircraft Recognition in High-Resolution Satellite Imagery

Abstract

High-resolution satellite images play an important role in defense monitoring, information gathering through reconnaissance and strategic intelligence systems. This research proposes a framework based on deep learning to automatically classify military aircraft using annotated satellite image datasets that contain horizontal bounding box data. The preprocessing steps for these datasets included removing undesired objects, cutting out selected images from the original image file, adding borders, enlarging images, reducing the depth of color in the pixels for each image so they were the same size and had the same amount of colors and creating one-hot encoded labels for each of the images. Three custom convolutional neural networks are developed using hyperparameter tuning with KerasTuner and Adam optimization, namely Custom CNN_1, Custom CNN_2 and Custom CNN_3. Different train-test split ratios are used, which were 90:10, 85:15, 80:20, 75:25, 70:30, 65:35 and 60:40. The final evaluation criteria for each of the models are accuracy, precision, recall and F1-score. Of the three Custom CNN models, Custom CNN_1 produced an average accuracy of 92.95% and had the highest overall accuracy for all of the experiments, while Custom CNN_3 had an individual maximum accuracy of 94.11%. The results of this research support that the custom CNN-based framework provides the necessary processes to accurately extract the spatial features that can discriminate between objects in high-resolution satellite imagery for the purposes of classifying military aircraft.