Medium-range Trajectory Prediction Network Compliant to Physical Constraint for Oceanic Eddy

This paper proposed a novel neural network structure to achieve medium-range eddy trajectory prediction, named ETPNet, which is compliant with the physical constraint. This structure contains a variant of the long short-term memory (LSTM) cell, enhancing the dynamic interaction and representation ability of the features, constraints, and knowledge. The pipeline of the whole prediction framework is shown as follows:

<p align="center"> <img src="img/model.png" alt="Centered Image" width="600" /> </p>

And the structure of the variant LSTM is illustrated as

<p align="center"> <img src="img/TraceLSTM.png" alt="Centered Image" width="600" /> </p>

Experimental Results

MAE and MSE are the traditional evaluation metrics for regression tasks in machine learning, which are used for quantitative analysis of the error between predicted and truth. In practical applications, the real distance needs to be calculated in conjunction with the characteristics of the earth. Thus, we adopt MGD and SGD, which are computed by the formula of great circle distance. The MGD is the mean geodesic distance and the SGD is the summation geodesic distance of the future seven days. | Method (MAGE Loss) | MGD | SGD | MSE | MAE | | :----------------- | :---------: | :---------: | :--------: | :--------: | | EGRU | 45.9186 | 321.4303 | 0.1964 | 0.2766 | | ELSTM | 32.1280 | 224.8962 | 0.5132 | 0.1818 | | Seq2Seq | 40.5100 | 283.5702 | 0.1219 | 0.2452 | | MSeq2Seq | 12.5271 | 87.6897 | 0.0124 | 0.0757 | | ETPNet-Ego | 10.4589 | 73.2123 | 0.0088 | 0.0634 | | ETPNet-Physical | 10.6959 | 74.8712 | 0.0090 | 0.0646 | | ETPNet | 10.1715 | 71.2008 | 0.0085 | 0.0611 |

| Method (L1Loss) | MGD | SGD | MSE | MAE | | :-------------- | :-----: | :------: | :----: | :----: | | EGRU | 58.4550 | 409.1852 | 0.3792 | 0.3311 | | ELSTM | 36.4317 | 255.0218 | 0.6370 | 0.2060 | | Seq2Seq | 37.8517 | 264.9622 | 0.1120 | 0.2155 | | MSeq2Seq | 14.2436 | 99.7053 | 0.0145 | 0.0848 | | ETPNet-Ego | 14.1674 | 99.1718 | 0.0169 | 0.0859 | | ETPNet-Physical | 18.1555 | 127.0886 | 0.0264 | 0.1094 | | ETPNet | 10.2154 | 71.5075 | 0.0086 | 0.0616 |

Citing

if you find this work is helpful to your research, please consider citing our paper.

@article{10190732,
  author={Ge, Linyao and Huang, Baoxiang and Chen, Xiaoyan and Chen, Ge},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Medium-Range Trajectory Prediction Network Compliant to Physical Constraint for Oceanic Eddy}, 
  year={2023},
  volume={61},
  number={},
  pages={1-14},
  doi={10.1109/TGRS.2023.3298020}