| Authors |
Abbas, Muhammad John ; Khan, Muhammad Attique ; Hamza, Ameer ; Alsenan, Shrooq ; Alasiry, Areej ; Marzougui, Mehrez ; Shin, Jungpil ; Nam, Yunyoung |
| Abstract [eng] |
Remote sensing has always been an area of interest for researchers due to its significance in Earth monitoring, which supports proper future planning for agriculture, construction, reforestation, and climate change. Transformer architecture achieves significant performance in remote sensing image classification; however, they come with the trade-off of higher computational complexity. In this paper, we propose a novel deep learning framework, DFSNet-VLM—Cross Domain Fusion based Texture-Sensitive Dual Stream Network — for high-precision remote sensing scene understanding. The proposed framework includes a classification model, “DFSNet,” that improves feature representation by employing both spatial and frequency domain features, which ultimately help detect global patterns and textures alongside local features. The model also promotes information exchange between both streams to complement one type of features with respect to the other by integrating cross-domain fusion blocks at multiple stages. Additionally, a pretrained VLM model, “BLIP-2,” is integrated to provide semantic descriptions of classified images. Bayesian optimization is applied to fine tune hyperparameters, reducing overfitting and improving model performance. The proposed model is evaluated on six diverse publicly available datasets and achieves improved accuracies of 97.13% on MLRSNet, 94.67% on NWPU-RESISC-45, 98.00% on EuroSAT, 92.25% on GeoSceneNet16k, 98.25% on cloud, and 96.03% on the Bijie-landslide dataset, respectively. Detailed ablation studies, comparative analysis, and Grad-CAM++-based model explainability demonstrate that the proposed model is generalizable and scalable, and that it achieves improved accuracy. In addition, the proposed model can be easily implemented in a real-time environment for diverse applications. The trained model's links are available in the data availability section. |
