The use of Uncrewed Aerial Vehicles (UAVs) for high-resolution Earth Observation is revolutionizing large-scale Solar Farms inspection. However, the critical bottleneck remains the real-time, onboard analysis of thermal data. This paper introduces EDGE-SFOS v1.0, a novel Geo EDGE-AI framework that transforms the UAV into an intelligent agent capable of autonomous, in-situ fault detection. The core scientific contribution is a definitive, real-world performance comparison of two state-of-the-art tiny models deployed on our embedded system. We evaluate a modern transformer-based model, RF-DETR, against a leading-edge convolutional neural network, YOLO 12. The results are conclusive. Deployed via the EDGE-SFOS platform, RF-DETR delivered superior performance, achieving a significantly higher detection accuracy (0.58 vs. 0.52 mAP) and proving to be 24% faster (4.96 ms vs. 6.13 ms inference time) than its YOLO 12 counterpart. This work establishes that for demanding Geo EDGE-AI tasks, modern transformer architectures can surpass top-tier convolutional models in both accuracy and speed on resource-constrained hardware, providing a validated blueprint for the next generation of intelligent field robotics.