Developing robust perception systems for autonomous Search and Rescue (SAR) robots remains a critical challenge due to the scarcity of annotated data from hazardous environments. Real-world datasets are constrained by safety limitations and rarely capture the visual complexity, structural variability, and sensor artifacts characteristic of disaster scenarios. This work introduces a reproducible methodology for generating high-fidelity RGB datasets using NVIDIA IsaacSim, enabling the photorealistic simulation of post-disaster environments with controllable terrain textures, occlusions, and dynamic sensor effects. The pipeline incorporates automatic pixel-wise semantic labeling and is used to train a brain-like neural network model called Bayesian Confidence Propagation Neural Network (BCPNN) that first learns representations in an unsupervised manner and then classifies terrain textures once the labels are made available. The proposed framework is validated through extensive experiments in both simulated environments and physical testbeds that recreate representative SAR conditions under controlled settings. Results show that BCPNN models trained exclusively on synthetic data can generalize effectively to real-world RGB inputs, capturing terrain semantics with sufficient fidelity to support autonomous mobility decisions. This contribution provides a scalable data generation and learning pipeline for perception in extreme environments and establishes a practical foundation for deploying probabilistic terrain understanding in field-ready SAR robotics. The datasets of this development are available in the Zenodo repository. and code in the GitHub repository.