We investigate the use of system-on-chip (SoC) field-programmable gate array (FPGA) for accelerating deep learning (DL) algorithms onboard small satellites. We present an artificial intelligence (AI)-driven edge computing (EC) framework for onboard cloud detection, image prioritization, and adaptive compression, using a deep learning processing unit (DPU) on a Zynq UltraScale+ MPSoC. The framework integrates a U-Net-based cloud segmentation, a prioritization algorithm, and a convolutional autoencoder (CAE)-based image compression, forming a hierarchical onboard data filtering pipeline. The DPU-accelerated U-Net achieves 97.46% accuracy, with an inference time of 29.57 ms per patch and a power consumption of 2.4 W. The CAE compression model reduces data volume by 4× while maintaining high image fidelity (SSIM: 0.957, PSNR: 34.86 dB), with low inference time (5.73 ms) and power consumption (2.5 W). Experimental results confirm that the FPGA-accelerated AI framework enables low-latency, low-power, and scalable processing, optimizing real-time, autonomous decision-making for missions and reducing bandwidth utilization through selective data acquisition and transmission. This study highlights the transformative potential of FPGA hardware accelerators for enabling low-latency and low-power onboard data processing in resource-constrained space missions. The integration of real-time autonomous operations enhances mission decision-making, adaptive planning, and efficient data management, addressing critical challenges in modern Earth observation and emerging space applications.