DEAF: An adaptive feature aggregation model for predicting soil CO2 flux

Abstract

Soil-derived CO2 flux (SCF) plays a key role in mitigating global warming, and the accurate prediction of SCF is the basis for the accurate prediction of climate change. To explore the changes in SCF, discover soil respiration patterns, and accurately predict soil respiration CO2 flux, we used a self-developed soil respiration monitoring device to collect soil CO2 flux data, and proposed a Dish-ECA-Adain-Autoformer (DEAF) combination model with adaptive feature aggregation and bias correction of the Autoformer improvement network to accurately predict the changes in soil respiration. To explore the potential feature patterns in the CO2 flux time series data, the trend and seasonal features of the data were enhanced using inverse contrast based on Fourier continuous transform. The irregular features of the local segments in the unit time were eliminated by adaptive aggregation convolution, and the residual linkage was applied to obtain the potential paradigm of the sequence changes. In the self-constructed CO2 flux dataset, MSE and MAE of the proposed model reached 0.296 and 0.328, respectively, with an average improvement of 34.4% compared to the unimproved model. In the experiments on the public time series dataset, the average improvement was approximately 20.2%, indicating that the model can effectively mine the time series information of the data and the correlation information of the feature dimensions to improve the prediction performance and verify the accuracy and universality of the proposed model. Additionally, the short-, medium-, and long-term prediction results showed that the DEAF model fitted the soil CO2 flux data well, providing a reference basis for the accurate prediction of soil CO2 flux.

Keywords

• Soil-derived CO2 flux
• Time series prediction
• Deep learning
• Attention mechanism