<p>The field of multispectral (MS) image classification has attracted a lot of attention from the research community. One of the key challenges in hyperspectral image classification is to obtain high accuracy and high classification performance in spectral-spatial domains. Conventional schemes, including the Bowerbird (CSBO-KELM) and the Genetic Algorithm Multilayer Perceptron (GAMLP), have been developed and often perform poorly because of their limited ability to address the issues of complex spectral variability. To this end, the present work conveys an automated parameter tuned DL enabled MS image classification system. The real-world analysis on satellite data using an enhanced technique called Multi-modal Glow worm Optimizer (MGO) with Optimized Residual Spectral-Spatial Deep Network (ORSSDN) is the main emphasis of the submitted work. A spatial attention module called Advanced Principal Component Analysis (APCA) is designed for spatial-guided dimensionality reduction to avoid redundant features. MGO does the job of spectral band selection and hyperparameter tuning, while the Adam optimizer updates network weights. A residual block incorporates a sequential spectral-spatial consideration module to prevent overfitting and expedite the training process of the proposed model. The convergence stability of the system has been improved by a hybrid global-local optimization method. The performance of the work has been analysed using Gaofen-2 dataset and it is observed that improved classification performance has been attained with maximum accuracy of 96.5, F-score of 0.949, and MCC of 0.862. In addition to this, the robustness of the proposed model has been analysed in the adverse conditions of atmospheric noise and seasonal variations. Still broader analysis with multi-sensor and real seasonal datasets is left for future analysis.</p>