PaperlayerView exact revision on arXiv

Source-linked AI summary

Prediction of Sea Surface Temperature using Long Short-Term Memory

Qin Zhang, Hui Wang, Junyu Dong, Guoqiang Zhong, Xin Sun

arXiv:1705.06861v1cs.CV

TL;DR

Accurate SST prediction is difficult, especially in uncertain coastal seas, and existing approaches include computationally demanding physics-based models and data-driven methods. The paper formulates SST forecasting as time-series regression and combines LSTM with a fully connected layer. Experiments report effective coastal-sea prediction, including one-week and one-month forecasts, with online updating improving performance.

  • Problem

    SST prediction is important for climate and ocean applications but remains difficult because of uncertainty, especially in coastal seas.

  • Method

    The paper formulates SST prediction as time-series regression and combines an LSTM layer with a fully connected layer.

  • Results

    The LSTM network achieves the best prediction performance in the reported comparison, and the updated model performs best for 2016 prediction.

  • Takeaways & Limitations

    The proposed method demonstrates effective SST prediction for coastal seas of China and supports online model updating.

Abstract

from arXiv · show

This letter adopts long short-term memory(LSTM) to predict sea surface temperature(SST), which is the first attempt, to our knowledge, to use recurrent neural network to solve the problem of SST prediction, and to make one week and one month daily prediction. We formulate the SST prediction problem as a time series regression problem. LSTM is a special kind of recurrent neural network, which introduces gate mechanism into vanilla RNN to prevent the vanished or exploding gradient problem. It has strong ability to model the temporal relationship of time series data and can handle the long-term dependency problem well. The proposed network architecture is composed of two kinds of layers: LSTM layer and full-connected dense layer. LSTM layer is utilized to model the time series relationship. Full-connected layer is utilized to map the output of LSTM layer to a final prediction. We explore the optimal setting of this architecture by experiments and report the accuracy of coastal seas of China to confirm the effectiveness of the proposed method. In addition, we also show its online updated characteristics.

I. INTRODUCTION

SST prediction matters for climate, ocean, and coastal applications, yet uncertainty makes coastal-sea accuracy difficult. The paper uses LSTM-based data-driven modeling to predict SST and reports results for coastal seas of China.

  • SST is important for Earth’s surface energy balance, ocean heat, climate, and biological systems.
  • SST prediction supports ocean-weather and climate forecasting, offshore activities, environmental protection, and military applications.
  • Physics-based models use differential equations but are sophisticated, computationally demanding, and dependent on sea area.
  • Data-driven approaches learn SST models from observations using methods including linear regression, support vector machines, and neural networks.
  • The proposed network combines LSTM and a fully connected layer to model temporal relationships and produce SST predictions.
  • Experiments focus on fluctuating coastal seas of China and report prediction accuracy beyond existing methods.

II. METHODOLOGY

The paper formulates SST prediction as time-series regression and uses LSTM to capture temporal dependencies. LSTM’s gated memory regulates information flow to address long-term sequence modeling.

  • A. Problem formulation: SST values at one spatial grid form a time series, enabling prediction of future values from historical observations.
  • A. Problem formulation: The regression task maps k days of SST observations to predictions for days k + 1 through k + l.
  • B. Long short-term memory: LSTM is adopted to capture temporal relationships among SST time-series data.
  • B. Long short-term memory: LSTM processes input and hidden-state sequences while producing estimated outputs, new hidden vectors, and new memory vectors.
  • B. Long short-term memory: The cell state stores long-term information, while input, forget, output, and control gates regulate information entering, leaving, or being forgotten.

C. Basic LSTM blocks

A basic LSTM block combines an LSTM layer with a fully connected layer for single-grid SST prediction. The latter transforms the final hidden vector into the prediction.

  • The basic block combines an LSTM layer and a fully connected layer.
  • The LSTM layer captures temporal variation among SST time-series values.
  • The fully connected layer abstracts and combines the final hidden vector, reduces its dimensionality, and maps it to a final prediction.
  • A basic block predicts future SST for one grid from that grid’s historical SST values, while area prediction requires multiple blocks.

D. Network architecture

The network organizes spatial and temporal SST data as a cuboid architecture. Latitude and longitude define spatial dimensions, while time forms the third dimension.

  • The network architecture is represented as a cuboid whose axes are latitude, longitude, and time.
  • Grids at the same spatial location across the time axis form a basic LSTM block.

A. Study area and data

The study uses NOAA daily SST data and focuses evaluation on the more variable coastal seas near China, specifically the Bohai Sea. The Bohai subset forms a 16 by 15 grid for area prediction.

  • Data source: NOAA SST data contain 12,868 daily values from September 1981 to November 2016 on a 0.25-degree global grid.The grid covers 1440 × 720 latitude-longitude cells.
  • Study area: The evaluation focuses on coastal seas near China because their temperatures fluctuate more than those in the far ocean.The selected region is the Bohai Sea.
  • Network context: The network figures depict the LSTM block, full-connected layer, and overall network architecture used for the prediction model.These figures provide structural illustrations rather than additional study-area or data descriptions.
  • Study area: The Bohai SST Dataset is a 16 by 15 subset spanning 37.07N–41N and 117.35E–121.10E.It contains the full 12,868-day record for the corresponding Bohai region.

B. Experimental Setup

The experiments formulate SST forecasting as sequence prediction, use historical observations to predict future values, and evaluate one-week and one-month horizons. Training uses Adagrad with fixed data splits and RMSE/ACC metrics.

  • Problem setup: The model uses about four times the prediction length as historical input, balancing expected prediction quality against computational cost.This choice follows the periodic characteristics of the temperature data.
  • Experimental sequence: The study first determines critical network parameters at a single position, then evaluates area SST prediction for the Bohai Sea.The parameter experiment uses a basic LSTM block before the area-level evaluation.
  • Optimization: Adagrad is used with an initial learning rate of 0.1 and batch size 100 for network training.Adagrad adapts learning rates by parameter frequency.
  • Data splits: Data from 1981–2012.8 form training, 2012.9–2012.10 validation, and 2013–2015 testing, with 2016 reserved for comparison.The evaluation covers seven-day and 30-day forecasts and includes SVR for comparison.
  • Metrics: RMSE is minimized and ACC maximized; area prediction uses area average RMSE and area average ACC.RMSE and ACC are treated as absolute-error and relative-error measures, respectively.

C. Determination of parameters

The authors tune recurrent-layer depth, recurrent hidden units, and full-connected-layer depth using five Bohai positions. The selected configuration favors a single recurrent and full-connected layer with six recurrent units.

  • Recurrent units: Six recurrent units provide the most frequent best performance across the five positions, so the following experiments set units r = 6.The best value is observed most often, although positions p4 and p5 favor seven and five units, respectively.
  • Recurrent depth: The best recurrent-layer setting is lr = 1, possibly because additional layers increase weights beyond what the available data can train effectively.The experiment compares lr values from 1 to 3.
  • Recurrent depth: Table II reports RMSE and ACC for five positions across different lr settings.The accompanying experiment identifies lr = 1 as the best setting.
  • Full-connected depth: Table III reports RMSE and ACC for five positions across different k settings.The supplied passages describe the selected full-connected configuration but do not provide the table values.
  • Full-connected depth: The full-connected-layer experiment selects lfc = 1, with hidden units set equal to the prediction length.The passage attributes this choice to fewer trainable weights and lower computation.

D. Results and Analysis

The proposed LSTM method is compared with SVR using an RBF kernel on the Bohai SST dataset. The reported results identify LSTM as having the best prediction performance, with figures showing one-position predictions over one year.

  • Comparison: The experiment compares LSTM against SVR with an RBF kernel on the Bohai SST dataset.LSTM uses k values of 10, 15, 30, and 120 for prediction lengths of 1, 3, 7, and 30, respectively.
  • Results: The LSTM network achieves the best prediction performance according to area average RMSE and area average ACC.The table defines best performance as the smallest area average RMSE and largest area average ACC.
  • Visualization: Figure 5 compares one year of true SST values with predictions from LSTM and SVR at one position.The true series is green, LSTM predictions red, and SVR predictions blue.

E. Online model update

The online update experiment compares an original model with an updated model for 2016 SST prediction. Incorporating new observations improves the reported prediction performance and supports iterative updating without retraining from scratch.

  • E. Online model update: 2016 SST predictions compare the original model with an updated model trained by adding observations from 2013–2015.The 2016 data contain 328 days, and the results are reported in Table V.
  • E. Online model update: The updated model performs best on the 2016 Bohai Sea prediction results.Table V reports area average RMSE and ACC for the comparison.
  • E. Online model update: The proposed method supports online updating by feeding newly observed SST values back into the model during continued prediction.This procedure alternates prediction, observation collection, and model updating.
  • E. Online model update: Unlike SVR, the proposed method can incorporate new observations without retraining the regression model from scratch.The paper associates retraining from scratch with additional computing-resource costs.

IV. CONCLUSION

The paper formulates SST prediction as time-series regression and uses an LSTM network with a fully connected layer to predict future values. Experiments on coastal seas of China report prediction performance and online updating, while the network is described as independent of data resolution.

  • IV. CONCLUSION: The paper formulates SST prediction as time-series regression and uses LSTM to model temporal relationships for future-value prediction.The authors describe this as the first use, to their knowledge, of a recurrent neural network for SST prediction.
  • IV. CONCLUSION: The architecture combines an LSTM layer for time-series modeling with a fully connected layer that maps its output to a final prediction.The two-layer design is presented as the proposed SST regression model.
  • IV. CONCLUSION: The experiments report prediction performance for coastal seas of China and demonstrate online update characteristics.The conclusion identifies both evaluation scope and updating behavior as reported outcomes.
  • IV. CONCLUSION: The proposed network is described as independent of data resolution, with higher-resolution prediction requiring higher-resolution training data.The paper also mentions applications including ocean-front and abnormal-event prediction.
Loading 1705.06861v1…