TY  - JOUR
AU  - Farooq, Kawsar
AU  - Zarar, Muhammad
AU  - Muhammad, Yar
AU  - Hasnat, Muhammad
AU  - Li, Yanze
AU  - Luo, Xin
PY  - 2026
DA  - 2026/04/19
TI  - GeoFACT: Geospatial Few-Shot Adaptive and Compatible Training Framework for Remote Sensing Images Classification
JO  - ICCK Transactions on Applied Intelligence and Cybernetics
T2  - ICCK Transactions on Applied Intelligence and Cybernetics
JF  - ICCK Transactions on Applied Intelligence and Cybernetics
VL  - 1
IS  - 1
SP  - 53
EP  - 66
DO  - 10.62762/TAIC.2026.348838
UR  - https://www.icck.org/article/abs/TAIC.2026.348838
KW  - incremental learning
KW  - few-shot learning
KW  - remote sensing
KW  - Geo-FACT
KW  - image processing
AB  - Remote Sensing (RS) data streams necessitate classification models capable of assimilating novel information without the computational burden of retraining from scratch. Although Few-Shot Incremental Learning (FSIL) offers a promising paradigm for adaptation using limited samples, existing methodologies often falter due to the high inter-class similarity and complex background clutter inherent in aerial imagery. To address these challenges, we introduce a novel framework, Geospatial Few-Shot Adaptive Compatible Training, designed to mitigate catastrophic forgetting in dynamic environments. This method ensures forward compatibility by strategically allocating embedding space for future categories via virtual prototypes, while simultaneously maintaining backward compatibility through knowledge distillation and selective network freezing. The proposed framework was rigorously evaluated on the AID and NWPU-RESIS45 benchmark datasets to assess its ability to balance plasticity for new tasks with stability for previously learned classes. Experimental analysis using accuracy, loss, and F1-score metrics demonstrates the robustness of the approach, achieving a base-session accuracy of 96.62% on the AID dataset and 96.74% on the NWPU-RESISC45 dataset. These results confirm that the proposed model effectively addresses the limitations of current state-of-the-art techniques, offering a scalable solution for RS classification, Catastrophic Forgetting, Forward Compatibility, and Few-Shot Incremental Learning.
SN  - 3143-0309
PB  - Institute of Central Computation and Knowledge
LA  - English
ER  - 

@article{Farooq2026GeoFACT,
  author = {Kawsar Farooq and Muhammad Zarar and Yar Muhammad and Muhammad Hasnat and Yanze Li and Xin Luo},
  title = {GeoFACT: Geospatial Few-Shot Adaptive and Compatible Training Framework for Remote Sensing Images Classification},
  journal = {ICCK Transactions on Applied Intelligence and Cybernetics},
  year = {2026},
  volume = {1},
  number = {1},
  pages = {53-66},
  doi = {10.62762/TAIC.2026.348838},
  url = {https://www.icck.org/article/abs/TAIC.2026.348838},
  abstract = {Remote Sensing (RS) data streams necessitate classification models capable of assimilating novel information without the computational burden of retraining from scratch. Although Few-Shot Incremental Learning (FSIL) offers a promising paradigm for adaptation using limited samples, existing methodologies often falter due to the high inter-class similarity and complex background clutter inherent in aerial imagery. To address these challenges, we introduce a novel framework, Geospatial Few-Shot Adaptive Compatible Training, designed to mitigate catastrophic forgetting in dynamic environments. This method ensures forward compatibility by strategically allocating embedding space for future categories via virtual prototypes, while simultaneously maintaining backward compatibility through knowledge distillation and selective network freezing. The proposed framework was rigorously evaluated on the AID and NWPU-RESIS45 benchmark datasets to assess its ability to balance plasticity for new tasks with stability for previously learned classes. Experimental analysis using accuracy, loss, and F1-score metrics demonstrates the robustness of the approach, achieving a base-session accuracy of 96.62\% on the AID dataset and 96.74\% on the NWPU-RESISC45 dataset. These results confirm that the proposed model effectively addresses the limitations of current state-of-the-art techniques, offering a scalable solution for RS classification, Catastrophic Forgetting, Forward Compatibility, and Few-Shot Incremental Learning.},
  keywords = {incremental learning, few-shot learning, remote sensing, Geo-FACT, image processing},
  issn = {3143-0309},
  publisher = {Institute of Central Computation and Knowledge}
}