Estimating the ocean subsurface thermal structure (OSTS) based on multisource sea surface data in the western Pacific Ocean is of great significance for studying ocean dynamics and El Niño phenomenon, but it is challenging to accurately estimate the OSTS from sea surface parameters in the area. This paper proposed an improved neural network model to estimate the OSTS from 0–2000 m from multisource sea surface data including sea surface temperature (SST), sea surface salinity (SSS), sea surface height (SSH), and sea surface wind (SSW). In the model experiment, the rasterized monthly average data from 2005–2015 and 2016 were selected as the training and testing set, respectively. The results showed that the sea surface parameters selected in the paper had a positive effect on the estimation process, and the average RMSE value of the ocean subsurface temperature (OST) estimated by the proposed model was 0.55 °C. Moreover, there were pronounced seasonal variation signals in the upper layers (the upper 200 m), however, this signal gradually diminished with increasing depth. Compared with known estimation models such as the random forest (RF), the multiple linear regression (MLR), and the extreme gradient boosting (XGBoost), the proposed model outperformed these models under the data conditions of the paper. This research can provide an advanced artificial intelligence technique for estimating subsurface thermohaline structure in major sea areas.

Underwater object detection currently faces many challenges, such as the large number of parameters in existing object detection models, slow inference speed, blurring of underwater images, and aggregation of small targets, making it difficult to conduct efficient underwater object detection. This paper proposes a lightweight underwater object detection algorithm based on YOLOv5.The method uses depth-wise separable convolution instead of ordinary convolution to reduce the number of parameters and computational complexity. A C3 module based on Ghost convolution is designed to further compress the model size and improve the computational speed. In the feature extraction stage, a RepVgg module based on structural reparameterization is used to convert the multi -branch structure into a single-branch structure in the inference stage, improving the feature extraction ability of the model and increasing the inference speed. A Rep-ECA module is designed to embed the efficient channel attention module ECANet into the RepVGG module, selecting more effective channel information and improving the model’s feature extraction ability for small objects in blurred images, thereby improving detection precision. Experimental results show that in the URPC underwater object detection dataset, the proposed algorithm has a 39% lower model parameter count compared to the original model, a 42% reduction in computational complexity. The model can achieve a frame rate of 85 on a single Nvidia GTX 1080ti GPU, which is a 24% improvement over the original model, while mAP reaches 85.1%, a 1.1% improvement over the original model. The algorithm can improve the detection precision and achieve lightweight, which lays a foundation for the deployment of underwater equipment.