The Southern Ocean is a complex system yet is sparsely sampled in both space and time. These factors raise questions about the confidence in present sampling strategies and associated machine learning (ML) reconstructions. Previous studies have not yielded a clear understanding of the origin of uncertainties and biases for the reconstructions of the partial pressure of carbon dioxide (pCO(2)) at the surface ocean (pCO(2)(ocean)). We examine these questions through a series of semi-idealized observing system simulation experiments (OSSEs) using a high-resolution (+/- 10 km) coupled physical and biogeochemical model (NEMO-PISCES, Nucleus for European Modelling of the Ocean, Pelagic Interactions Scheme for Carbon and Ecosystem Studies). Here we choose 1 year of the model sub-domain of 10 degrees of latitude (40-50 degrees S) by 20 degrees of longitude (10 degrees W-10 degrees E). This domain is crossed by the sub-Antarctic front and thus includes both the sub-Antarctic zone and the polar frontal zone in the south-east Atlantic Ocean, which are the two most sampled sub-regions of the Southern Ocean. We show that while this sub-domain is small relative to the Southern Ocean scales, it is representative of the scales of variability we aim to examine. The OSSEs simulated the observational scales of pCO(2)(ocean) in ways that are comparable to existing ocean CO2 observing platforms (ships, Wave Gliders, carbon floats, Saildrones) in terms of their temporal sampling scales and not necessarily their spatial ones. The pCO(2) reconstructions were carried out using a two-member ensemble approach that consisted of two machine learning (ML) methods, (1) the feed-forward neural network and (2) the gradient boosting machines. The baseline data were from the ship-based simulations mimicking ship-based observations from the Surface Ocean CO2 Atlas (SOCAT). For each of the sampling-scale scenarios, we applied the two-member ensemble method to reconstruct the full sub-domain pCO(2)(ocean). The reconstruction skill was then assessed through a statistical comparison of reconstructed pCO(2) cean and the model domain mean. The analysis shows that uncertainties and biases for pCO(2)(ocean) reconstructions are very sensitive to both the spatial and the temporal scales of pCO(2) sampling in the model domain. The four key findings from our investigation are as follows: (1) improving ML-based pCO(2) reconstructions in the Southern Ocean requires simultaneous high-resolution observations (<3 d) of the seasonal cycle of the meridional gradients of pCO(2)(ocean); (2) Saildrones stand out as the optimal platforms to simultaneously address these requirements; (3) Wave Gliders with hourly/daily resolution in pseudomooring mode improve on carbon floats (10 d period), which suggests that sampling aliases from the 10 d sampling period might have a greater negative impact on their uncertainties, biases, and reconstruction means; and (4) the present seasonal sampling biases (towards summer) in SOCAT data in the Southern Ocean may be behind a significant winter bias in the reconstructed seasonal cycle of pCO(2)(ocean).