This chapter explores the impact of domain generalization methods and pruning strategies on model performance in two visual recognition tasks using the PACS benchmark and a custom Plant Type dataset. Using ResNet-18 as the backbone, we evaluate five domain generalization algorithms—Empirical Risk Minimisation (ERM), ERM++, Enhanced Quantile Risk Minimisation (EQRM), Invariant Risk Minimisation (IRM), and Uncertainty Risk Minimisation (URM)—under visual domain shifts: stylized domains (Photo, Art, Cartoon, Sketch) in PACS and degraded domains (blur, noise, low lighting) in plant images. We also assess three pruning techniques—unstructured, structured, and SNIP—at various sparsity levels to understand their impact on generalisation. Empirically, SNIP and unstructured pruning at light-to-moderate levels (10%–30%) improved accuracy by up to 4.2% in ERM and ERM++ on PACS, and by up to 2.8% in EQRM on the Plant dataset. Structured pruning, in contrast, caused accuracy drops of up to 7% in fine-grained plant classification tasks. Our findings reveal that domain generalisation techniques, when combined with effective pruning, can enhance model robustness across diverse environments. This approach enables models to maintain strong performance in real-world applications despite unseen or degraded input conditions, supporting the development of more reliable visual recognition systems.