Robust monocular depth estimation is pivotal for enabling autonomous UAVs to navigate cluttered and unstructured environments. Despite substantial progress using CNNs and Vision Transformers, existing methods often fail to adequately model long-range contextual relationships and multi-scale structural cues – limitations acutely detrimental to depth accuracy in complex aerial scenes critical for UAV operation. While existing methods leverage pre-trained encoders like VITs, Resnet, etc. for feature extraction, they often freeze these backbones, limiting their adaptability to depth-specific representations. In this work, an end-to-end trainable architecture that jointly optimizes a graph module composed of a DINOv2 encoder and a GraphSAGE layer, and a UNet backbone for monocular depth estimation is proposed and tested. The proposed architecture enables the model to learn depth-aware features while preserving rich semantic priors. Experiments on WHU aerial depth dataset highlight the benefits of training DINOv2 in conjunction with graph-based feature refinement, offering new insights into scalable and adaptive depth estimation. Ablation studies demonstrate the robustness and generalizability of the proposed model, particularly in complex scenes with varying depth structures.