Scene based classification in high resolution spatial satellite imagery is a challenging task due to intrinsic patterns. The extensive features in remotely sensed images rely on effectual image representation for scene understanding, thereby rendering precise classification. Over time, deep learning has become essential in advanced computer vision tasks, particularly in satellite imagery. Tuning hyper-parameters in deep learning remains a challenge due to model’s complexity, sensitivity towards learning features and the need to avoid overfitting. Towards this, we have proposed a deep learning based convolutional neural network (CNN) for classification of Land use/cover (LU/LC) in aerial images, named H-ADNET, with hyper-parameters optimization. The methodology employs a customized DenseNet121 CNN model with random search for hyper-parameters tuning to get the best fit model on EuroSAT aerial dataset with 10 classes. For better convergence of the proposed H-ADNET model, a wide range of simulations has been conducted by varying multiple parameters. A high performance classification accuracy of 95% has been achieved, demonstrating the proposed H-ADNET is capable to handing complex spatial features in aerial imagery.