This study aims to explore and implement drone-based human action recognition using deep learning architectures such as Res net, Dense net, CNN-LSTM, Efficient net, and Inception V3. The methodology elaborates on the deep learning models used, including YOLOV8, Open pose, Alpha pose, Resnet, Densenet, and Convolutional Neural Networks (CNNs) with Long Short Term Memory (LSTM), detailing their role in attribute recognition, pose estimation, and action classification. The system utilizes two widely used benchmark datasets: the UAV – human and UCF–101 datasets. Experimental evaluations demonstrate the system's robust performance across key tasks: attribute recognition with an average accuracy of 86% using the Dense net model; pose estimation with effective identification of key points in varied conditions through the Open pose method; and action recognition with an exceptional unprecedented accuracy of 99% using the EfficientNet-B0 model. These results underscore the system's ability to accurately identify a wide range of human activities in real - time, highlighting its potential applications in surveillance, security, disaster response, and other fields requiring dynamic and comprehensive aerial monitoring.