Automatic classification of remote sensing images for land use and land cover (LULC) applications encounter numerous challenges because of multi-resolution data, heterogeneous appearance and multi-spectral complexity. Existing deep learning models, especially conventional CNNs, struggle to include inherent features of objects because of their limited ability in modeling the long-range dependency and global contextual information. To address these limitations, this paper proposes a novel Heterogeneous Dual-Path Wide Residual Network (HDP-WRN) model for enhanced LULC classification. The initial feature extraction is performed using a post-activation residual block and two parallel paths using different convolutional operators with different attention mechanisms. Branch A integrates standard convolutions and Efficient Channel Attention (ECA) to extract fine-grained local textures, and Branch B merges Ghost convolutions and Convolutional Block Attention Module (CBAM) to boost global spatial channel feature extraction effectively and their complementary representations are fused for obtaining discriminative joint feature space effectively. Extensive experiments show that HDP-WRN achieves competitive performance on multiple benchmark datasets, with accuracy of 98.67% on EuroSAT, 95.11% on RSSCN, 96.13% on SIRI-WHU, and 97.24% on UC-Merced. Under our experimental setup, the proposed model outperforms the reported results of several existing methods, though direct comparisons should be interpreted with caution due to differences in training protocols across studies. Results verify the model for accurate and robust LULC classification for diverse spatial resolutions and remote sensing images.