Abstract: With the development of deep learning, image segmentation models have been increasingly applied to obstacle detection in variable geometry design for 3D seismic acquisition. Most exploration areas cover a large and irregular range, and the remote sensing images used for observation system design are usually large in size and contain a large amount of invalid data. Deep learning methods for obstacle recognition focus on improving accuracy but lack pre-recognition processing of remote sensing images and post-recognition procedures such as obstacle boundary vectorization, which significantly limits their applicability. To overcome these limitations, we propose a "four-step" method for large-scale remote sensing image processing: remote sensing images preprocessing, automatic obstacle recognition, postprocessing, and variable geometry design for obstacle avoidance. This end-to-end workflow enables automatic obstacle localization in large-scale remote sensing images and subsequent variable geometry design. First, an adaptive preprocessing method is formulated to fit the minimum circumscribed rectangle of the valid data region for optimized image tiling. Next, an enhanced UMiT-Net (Softmax) model is proposed to simultaneously extract different types of obstacles. The subsequent vectorization techniques incorporate boundary extraction, stitching, expansion, and coordinate transformation to derive projected coordinates. Finally, an obstacle-avoidance algorithm relocates shot and receiver points to determine optimized shot coordinates for geometry adjustment. Experimental results demonstrate that the proposed method eliminates 50.4% of invalid image tiles in a real-world exploration area, improving computational efficiency by 102%. Obstacle detection achieves high accuracy and completeness, and vectorized boundaries enable rapid obstacle-avoiding geometry design. The end-to-end workflow is both efficient and user-friendly.