Toward spectrally truthful models for gap-filling soil respiration and methane fluxes. A case study in coastal forested wetlands in North Carolina

•Energetic frequencies of drivers were identified using wavelet multiresolution analysis.•Those frequencies were used to gap-fill soil respiration and methane fluxes.•Wavelet multiresolution - machine learning fusion was highly efficient (r2∼ 0.9) in gapfillingfluxes•Nearly 70–90 % of measured data points were incorporated with 95 % credible interval with Bayesian approach.•Performance of the new approach was comparable to measured-data driven MDS approach. Soil respiration (Rs) and methane (FCH4) fluxes are two important metrics of ecosystem metabolism. An accurate estimate of the budget of these two greenhouse gases is critical to understanding their response to climate and land-use changes. Reconstructing continuous time series of gappy chamber Rs and eddy-covariance derived FCH4 measurements is usually done based on correlative relationships of these fluxes with environmental variables. However, current approaches do not account for the fact that different environmental drivers affect the carbon fluxes at different temporal scales. Here we propose a novel gapfilling technique that accounts for the specific spectral frequencies at which each of the environmental variables covaries with Rs and FCH4 - photosynthetically active radiation at diel scale, soil temperature at synoptic scale, and soil moisture, water table depth and atmospheric pressure at synoptic and seasonal scale. The method was applied on two operational loblolly pine plantations of different ages and a mixed hardwood forested wetland on the lower coastal plain of North Carolina. The time series of these environmental drivers were reconstructed using wavelet decomposition and a Daubechies wavelet filter. Further, to consider the joint influence of the environmental drivers, parametric (elastic net regression, support vector machine, gradient boost and artificial neural network), and nonparametric (Bayesian) statistical models were chosen, and compared the results with Q10 and Marginal Distribution Sampling (MDS) outputs. In all cases, the algorithms were trained on 70 % of the data and validated with the remaining data. Spectral-filtered models did not significantly differ from those driven by unfiltered data with respect to Rs and FCH4 predictions. While all the spectrally driven algorithms achieved high predictive accuracy against Q10, the increase in model fit compared to MDS was minimal. Spectral data filtering modestly improves model accuracy, shedding light on complex environmental a