Due to their small computational requirements, surrogate models are widely used to approximate complex systems that would otherwise require computationally expensive simulations. A key factor in establishing an accurate surrogate model is the design of the training dataset, which involves both the sampling strategy and the determination of the optimal number of samples. Traditional stationary sampling methods, such as Latin hypercube sampling, face limitations because the required number of samples is unknown a priori. This uncertainty can lead to additional overhead, especially in the case of highly complex numerical models where each simulation may require hours or even days to complete. To overcome this drawback, sequential sampling methods, and in particular adaptive sampling, are preferred, as they can increase the number of sampling points in regions of greater interest, with the aim of obtaining a better-performing surrogate model with fewer samples. This study introduces a cross-validation-based, model-independent adaptive sampling approach that employs an additional surrogate model to assess the error in unsampled regions. The proposed method is based on the Space-Filling Cross-Validation Tradeoff method proposed in [1] while considering an updated distance criterion and supporting batch selection to improve sampling efficiency and expedite enhancements in model quality. The performance of the proposed method is evaluated against two other sampling techniques: a Latin hypercube sampling with centred L2-discrepancy and the original method of [1]. Learning curve results derived from two 2D analytical functions serve as the basis for evaluation of the surrogate model obtained using Support Vector Regression [2]. Furthermore, the adaptive sampling method is applied to develop a high-dimensional surrogate model for a geotechnical application involving an urban excavation in London, demonstrating its capability to address complex problems.