Hybrid Learning with Multi-Objective Optimization for Remote Sensing Few Shot Scene Classification

Few-shot learning has important applications in remote sensing image classification due to its strong generalizability. The objective of few-shot learning is to leverage prior knowledge to address new tasks with only a small number of fine-tuning samples. Therefore, How to accumulate prior knowledge during training stage is a key problem in solving few-shot classification problem. In addition, most existing methods suffer from the class confusion because of the inter-class similarity and intraclass diversity of remote sensing images. To alleviate these problems, this paper proposes Hybrid Learning with MultiObjective Optimization (HL-MOO) to enhance few-shot remote sensing scene classification performance. Specifically, in training stage, we propose a joint learning strategy that combines two selfsupervised learning tasks and a classification task to capture the intrinsic features of remote sensing images. We also introduce a multi-objective optimization algorithm to dynamically adjust the impact of each task on our model. HL-MOO not only improves the performance on training samples, but also enhance adaptability to new scenes. It is evaluated on NWPU-RESISC45, Aerial Image Dataset (AID), and WHU-RS19 benchmarks, achieving average accuracies of $\mathbf{8 1. 0 9 \%, ~} \mathbf{7 8. 4 0 \%}$, and $\mathbf{9 6. 3 4 \%}$ in the 1-shot setting, and $89.58 \%, 88.12 \%$, and 97.89 % in the 5 -shot setting, respectively. We also test its cross-domain performance on UCM dataset. Results show that our HL-MOO exhibits improvements in few-shot remote sensing scene classification tasks compared to state-of-the-art models.