<p>The work presented in this paper aims to demonstrate the usefulness of multivariate variational mode decomposition (MVMD) based image feature maps in developing a hasta image classification system within the context of Bharatanatyam, a classical Indian dance form. The research emphasizes the importance of capturing variations across all hasta classes using MVMD for building a Convolutional Neural Network (CNN) -based hasta recognition system. Specifically, multivariate Intrinsic Mode Functions (IMFs) are extracted by applying MVMD to a stacked dataset of all 29 hasta classes, represented as a multi-channel data matrix. A class-independent hasta image is then generated by aggregating the first IMF modes of all classes. For each class, training and testing images are obtained by projecting each raw hasta image onto the eigenvector that shows the maximum class variation relative to the class-independent MVMD hasta image. The average of the scores obtained from CNN models trained on MVMD-projected, eigen-projected, and raw hasta images shows a significant improvement in recognition performance, highlighting the complementary class information captured by the different models. Additionally, experiments using pretrained networks such as MobileNetV1 and MobileNetV2 demonstrated a similar trend, with average score-level fusion across the three models further improving hasta recognition performance.</p>