This paper addresses a critical issue in seabed characterization with deep learning semantic segmentation using high-resolution Synthetic Aperture Sonar (SAS) data, that we call Catastrophic Receptive Field Overflow (CRFO). We propose novel methods, including Mosaic Augmentation and Homogeneous Patch Rejection, to (1) effectively mitigate CRFO and (2) enhance model performance. Through experiments on real-world SAS data, we investigate the origins of CRFO, revealing its dependence on model architectures and data characteristics. The presented solutions exhibit promising results, whether measured in terms of Overall Accuracy or the reliability of models in inference across various image input sizes or aspect ratios, in the face of new proposed metrics. These findings provide valuable insights for addressing CRFO challenges in tasks involving relatively homogeneous datasets.