<p>Knowledge of oceanic thermal structure is significant to characterize the stratification and mixing processes that influences the climate system and oceanic ecosystems. Though <i>in-situ</i> sensors are able to acquire accurate vertical temperature profiles, they are mostly limited to fixed locations and sparse extent. Reconstruction of vertical temperature profiles based on remotely sensed sea surface variables has become popular given its cost-effective at extended spatial coverage. To address this, we develop a data-driven model that reconstructs vertical temperature profiles in the Northwest Pacific Ocean using remotely sensed sea surface variables—specifically, sea surface temperature (SST) and sea level anomaly (SLA). The model architecture consists of three components: a convolutional neural network (CNN) to extract spatial features from a 21 × 21 SST matrix centered on each Argo profile, followed by two fully connected neural networks (NN₁ and NN₂) that integrate local SST and SLA to predict the vertical temperature profile. The trained model exhibits high accuracy and robustness in capturing essential vertical temperature features across various depths, as indicated by the high coefficient of determination and minimal error metrics. Given the high correlation of temperature with historical variables, an extended model with input up to 11 d prior to array for real-time geostrophic oceanography (Argo) observations is also evaluated. Its degraded performance suggests that the more recent data adds the most value in the vertical temperature profile reconstruction. Future work will focus on validating the model across different oceanographic regions and incorporating additional sea surface variables to enhance its predictive accuracy and applicability.</p>