A Novel Dense Generative Net Based on Satellite Remote Sensing Images for Vehicle Classification Under Foggy Weather Conditions

• Published in: IEEE transactions on geoscience and remote sensing 2023, Vol.61, p.1-10
• Main Authors: Yuan, Jianjun, Liu, Tong, Xia, Haobo, Zou, Xu
• Format: Article
• Language: English
• Subjects: Ablation; Classification; Convolutional neural networks; Deep learning; Dense Generative Net (DGNet); Environmental conditions; Feature extraction; Foggy weather conditions; High resolution; Image classification; Image enhancement; Image resolution; Intelligent transportation systems; Kernel; Machine learning; Meteorology; Neural networks; Receptive field; Recognition; Remote sensing; Satellite imagery; satellite remote sensing; Satellites; Task analysis; Transformers; Transportation networks; vehicle classification; Weather
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2023.3336546

Accurate vehicle-type classification plays a crucial role in the development of intelligent transportation systems. Recently, several deep learning models have been proposed to utilize satellite remote sensing images for vehicle-type classification. However, conventional neural network models often have limitations when dealing with remote sensing images, such as adverse weather conditions, as well as the extremely low resolution of remote sensing images containing small objects like vehicles. To enhance the vehicle-type classification capability in complex environments, this article develops a novel deep learning framework called Dense Generative Net (DGNet). DGNet consists of three components: feature layer, generation layer, and dense feature fusion layer. The feature layer employs large convolutions to establish a broader receptive field, enabling it to capture more effective global feature information. The generation layer is based on a super-resolution (SR) network, which is designed to generate high-resolution feature information. The dense feature fusion layer performs the final classification by integrating the outputs from two upstream branches and combines the feature information obtained from the feature layer and the generated high-resolution features information from the generation layer, enabling comprehensive and robust classification of vehicle types. To evaluate recognition capability, vehicle data from multiple regions and diverse environmental conditions are utilized, including four different weather conditions. The experimental results demonstrate that DGNet exhibits remarkable vehicle-type recognition capability, with minimal degradation even under heavy foggy weather conditions. The effectiveness of each module and its impact on the overall performance have been verified through ablation experiments.