The rapid expansion of aerial Internet of Things (IoT) applications—including drone-based surveillance, environmental monitoring, and disaster response—demands lightweight and energy-efficient deep learning solutions. However, resource-constrained aerial platforms present significant challenges for deploying computationally intensive models like Convolutional Neural Networks (CNNs) and Autoencoders (AEs). This paper proposes an efficient deep learning acceleration framework optimized for aerial IoT systems, built on the RISC-V architecture. The proposed system optimizes matrix multiplication operations and parallel computation to enhance performance. Implemented on the Zedboard platform, our framework achieves 97.16% accuracy while utilizing minimal hardware resources. Performance evaluations show a 30.5 \(\times \) speedup for convolutional layers and a 10 \(\times \) improvement for fully connected layers. Comparative analysis with TensorFlow, Google Colab, and Zynq-7000 highlights the system’s computational efficiency, achieving a 1.68 \(\times \) to 7.84 \(\times \) acceleration. These results demonstrate the potential of RISC-V architectures to enable real-time, energy-efficient AI inference for aerial IoT applications.