TNeXt: A convolutional neural network for remote harbor classification

Abstract

Harbor recognition from aerial images faces two main challenges: deep models are often too large for real-time use on drones, and there is no large, diverse harbor dataset to train them. This work overcomes these obstacles in two ways. We built the Turkish Harbor Image Dataset (THID) by flying a UAV over 207 harbors in T & uuml;rkiye and capturing 13,199 clear-weather images. We split THID into 73.4 % training, 18.3 % validation, and 8.3 % test sets, and applied simple augmentations (rotations, flips) to improve robustness. TNeXt, a fully convolutional network is proposed in this research. On THID, TNeXt achieved 97.71 % accuracy. Without changing its architecture, it scored 83.30 % top-1 on ImageNet1k. For the UC-Merced Land Use dataset, TNeXt reached 97.14 % accuracy; when used as a feature extractor in a simple pipeline, it hit 99.76 %. This research provides high accuracy and rapid inference and is therefore suitable for real-time harbor detection for autonomous platforms.