Abstract: As an important mesoscale ocean phenomenon, ocean temperature fronts are a key factor affecting ocean heat exchange, material transport and sea-air interaction. Accurate detection of ocean temperature fronts is crucial for analyzing temporal and spatial changes of ocean temperature fronts and dynamic monitoring of marine meteorological. Due to the ocean mixing and slow temperature change, the ocean temperature front has the characteristics of small targets and weak edges. In view of the problems of inaccurate morphological description and pixel misdetection in traditional edge detection methods and existing deep learning model, this paper proposes a multiscale ocean temperature front detection method, M-PSPNet. The ability of the model to detect edge contours and positional information is improved by designing the multiscale feature extraction module (Multi-ResNet), which retains the spatial and positional features obtained in the shallow learning network while combining them with the semantic features obtained in the deep network. Additionally, the hybrid loss function DF_loss, which combines Dice_loss and Focal_loss, is introduced to guide the model to focus on the pixel-level difference between the predicted result and the labeled value, improving the accuracy of frontal pixel detection. In order to verify the effectiveness of the proposed method, multiple groups of comparative experiments were designed based on the experimental model. The experimental results show that the M-PSPNet achieves 78.79%, 89.59%, 86.95%, and 88.25% respectively in IOU, Recall, Precision and F1 score. The model detection results using the Multi-ResNet module increased by 14.78%, 19.15%, and 10.13% in IOU, Recall, and F1 score. The model detection results using the DF_loss increased by 1.4%, 1.55%, and 5.1% in IOU, Recall and Precision. The comparison results demonstrate that the M-PSPNet is capable of accurately locate the position and edge contour of the ocean temperature front, and describe the accurate front shape.