Optical image sensing plays a major role in both military and civil sectors includes earth observation, target detection and surveillance. Generating captions for these images is a typical task, traditional methods are time-consuming, failed to handle multimodal image data, difficulty in handling complex patterns and prone to human error. Also, less scalability issues and failed to address the interpretability problems. To address these challenges, cutting-edge Machine and Deep learning models are used for automated analysis to improve the decision-making process and detection accuracy. Hence, to enhance the robustness of the detection method, an Attention-based Caption Generation approach for Optical Remote Sensing Images is proposed. It integrates a pre-trained VGG16 as an EncoderCNN process to extract high-level features and a multi-head attention mechanism to selectively emphases on extracted image regions. Finally, LSTM is used as a DecoderLSTM to generate captions word-by-word. The proposed approach improves the accuracy and robustness of caption generation by combining the spatial information from image features and sequential dependencies captured by LSTM method. The efficacy of the proposed approach is validated using the real-world optical image sensing dataset in terms of BLEU, METEOR, and ROUGE_L. The proposed model process 7 tokens and 4.25% improvement in BLEU scores with beam size 1 to 4.