Hyperspectral image (HSI) classification requires models that leverage long-range spectral and spatial dependencies while handling scarce labels and the high dimensionality of the data. This paper introduces a semi-supervised Graph Convolutional Network (GCN) that builds a graph over both labeled and unlabeled pixels to better capture the data manifold. It also proposes the implementation of Mini-GCN, an inductive mini-batch variant that preserves graph reasoning with far lower memory and computational cost. On top of these, a hybrid end-to-end FuNet architecture fuses a CNN with Mini-GCN (for mid- and long-range topology) via three fusion schemes: additive, multiplicative, and concatenation, to learn complementary representations. Experiments on two benchmark datasets, Indian Pines and Pavia University, using limited labeled training samples, are compared against various conventional and Deep Learning (DL) CNN-based algorithms, and also supervised GCN. Semi-supervised GCN already outperforms CNNs and supervised GCN; Mini-GCN further enhances efficiency without compromising accuracy, and the proposed fusion networks yield the best performance. Notably, FuNet-C attains an OA of 95.79% and κ of 0.95 on Indian Pines, and OA 92.36% and κ 0.90 on Pavia University, with marked gains on minority classes, confirming that combining mini-batch graph reasoning with CNN features is an effective, label-efficient paradigm for HSI classification.