Deep learning approach characterizes information by separating significant features automatically. Transfer learning approach for feature extraction from a pre-trained deep model is a well-known strategy in classification or target recognition task. In this paper, VGG-16, Inception-v3, and ResNet- 50 pre-trained models are utilized for extraction of features. Features are extracted from the patches of different locales like urban, forest and water, from the RGB image of Synthetic Aperture Radar data of RISAT-1. Extracted features are further trained on SVM for better classification. All the three networks are also re-trained by fine-tuning the fully connected layers to classify the patches. Among the three pre-trained models, VGG- 16 outperformed the other classifiers in terms of accuracy of classification and feature extraction.