Hyperspectral images (HSIs) are contiguous spectral information with prominent applicability in remote sensing. Classification of HSIs is a significant challenge because of their higher dimensionality and limited ground truth availability. Most of the CNN (Convolutional Neural Networks) based classification techniques pre-apply Dimensionality Reduction (DR) to deal with the curse of higher dimensionality, and hence, they failed to learn the spatial-spectral characteristics of the HSIs. The proposed T-HyC model is superior in accommodating spatial-spectral features and handling limited ground truth problems. The advancement of Transfer Learning (TL) provided a distinct advantage in addressing limited-label issues of HSIs. The proposed T-HyC operates in two stages: Multi-Scale 3D CNN (M3D-CNN) and TL based Spatial-Spectral CNN (SS-CNN). The M3D-CNN module has multiscale 3D-CNN blocks that extract spatial-spectral information at multiple scales. Since there are no such pre-trained models specifically for HSIs, we apply the learned representation of M3D-CNN for transferring knowledge to the classification block of SS-CNN. The SS-CNN is a series of 3D-CNN and 2D-CNN blocks that extract spatial and spectral features. T-HyC demonstrated superior classification performance even in conditions of limited training samples. This work substantiates the capability of CNN models in transferring knowledge gained from fine-labelled scenes to coarse-labelled scenes for classification. This work substantiates the capability of CNN models in transferring knowledge gained from fine-labelled scenes to coarse-labelled scenes for classification.