| Title |
LiteDenseMoE: an explainable lightweight densely connected mixture-of-experts network for aerial scene recognition in low contrast remote sensing images |
| Authors |
Abbas, Muhammad John ; Khan, Muhammad Attique ; Hamza, Ameer ; Alsenan, Shrooq ; Alasiry, Areej ; Marzougui, Mehrez ; Shin, Jungpil ; Nam, Yunyoung |
| DOI |
10.1109/JSTARS.2025.3645113 |
| Full Text |
|
| Is Part of |
IEEE Journal of selected topics in applied Earth observations and remote sensing.. Piscataway, NJ : IEEE. 2026, Early access, p. 1-20.. ISSN 1939-1404. eISSN 2151-1535 |
| Keywords [eng] |
Remote sensing ; Aerial scene ; Deep learning ; Hyperparameter tuning ; Mixture of Experts ; Model interpretations |
| Abstract [eng] |
Land Remote sensing image classification is crucial for understanding ongoing geographical and environmental changes. It aids in land use and land cover classification, crop and vegetation classification, change detection, and classification of coastal and aerial regions. Many advanced techniques were introduced based on some substantial modifications in the models; however, this resulted in a complex framework that is difficult to adapt. In this work, we proposed a novel Lightweight Dense Mixture of Experts (LiteDenseMoe) model for aerial and coastal regions classification using remote sensing images. The proposed model initially incorporates light, dense blocks with lightweight dense layers, as well as channel and spatial attention mechanisms. The resulting model is further fused with an Mixture of Experts block that extracted more relevant and essential features for the accurate prediction of complex aerial scenes. In the training process of the proposed model, a Hyperband Optimization technique is employed for hyperparameter initialization, rather than manual selection. After training the proposed model, classification was performed, along with output interpretation. The proposed LiteDenseMoe architecture is evaluated on three datasets and achieved an accuracy of 93.25% on MLRSNet, 92.56% on NWPU-RESISC45, and 96.54% on the EuroSAT dataset with only 0.3 million parameters. Expert allocation and their confidence per class, Expert disagreement Network, and t-SNE visualization are also observed to interpret the Moe results. Detailed Ablation studies and comparative analysis with pre-trained and SOTA models confirm the impact and efficiency of the proposed architecture for aerial and coastal regions classification. |
| Published |
Piscataway, NJ : IEEE |
| Type |
Journal article |
| Language |
English |
| Publication date |
2026 |
| CC license |
|
