Conventional transfer learning approaches involve transferring knowledge from a rich source domain to improve the performance of the target classifier in a relatively scarcer target domain. However, the models developed in the target domain are relatively lightweight and struggle with overfitting and generalization issues; therefore, their usability is negatively affected in relatively scarcer and noisy environments in domains like remote sensing. In this work, we propose a novel nuclear norm-induced lightweight and robust teacher-student network to classify scenes in the remote sensing domain. Specifically, the student model of the proposed teacher-student network is regularized using a nuclear norm constraint to obtain low-ranked student model parameters that are robust to noise in the environment. The student model of the proposed teacher-student network is trained on different benchmarked remote sensing datasets to classify scenes and subsequently used for inference. With the UC Merced and EuroSAT datasets, the classifier’s performance reaches up to \(87.0 \%\) . Moreover, the proposed teacher-student network parameters are low-ranked and robust to noise in the deployed environments. Hence, the proposed lightweight and robust nuclear norm-induced teacher-student network can be deployed to classify scenes using remote-sensing images in noisy environments.