11th International Conference on ICTIS

Authors - Yaram Srinivasa Reddy, Bairoju Sreelatha, Shankar Lingam. M
Abstract - Knowledge from a resource-rich source domain is leveraged in traditional transfer learning to enhance classification in a relatively data-scarce target domain. However, the resulting target models often suffer from overfitting and limited generalization, which restricts their utility in noisy and resource-constrained environments such as remote sensing. To mitigate these limitations, this work introduces a nuclear norm–regularized teacher–student framework for hyperspectral scene classification. In particular, the student model is regularized with the nuclear norm to encourage low-rank parameter representations, improving robustness to ambient noise. Further, we introduce a relative reconstruction loss (RRL) metric to measure the robustness of the student model to environment noise. Trained on several benchmark datasets, the proposed student model attains up to 87.0% classification accuracy on the independent test sets of UC Merced and EuroSAT, while remaining substantially lighter than the teacher network. Further, relative reconstruction values are computed for different amounts of noise added in the input space; RRL saturate to values less than 1.0 for all the datasets, substantiating that the regularized student model is indeed robust. The competitive performance of the regularized student model compared to the teacher network, its lightweight design, together with RRL values less than one, suggest that the proposed student model can effectively be deployed in noisy and resource-constrained environments such as edge and fog devices.