Prediction of carbon dioxide emissions from Atlantic Canadian potato fields using advanced hybridized machine learning algorithms – Nexus of field data and modelling

· CO 2 , soil characteristics, and climatic variables collected at three fields in Canada. · Three new ensemble ML models of AR-RF, MS-RF, and ICO-AR-RF were developed for CO 2 prediction. · Air temperature, dew point, and soil temperature are the three most effective variables on CO 2 . · ICO-AR-RF model performed the best, followed by MS-RF and AR-RF. · Uncertainty analysis showed that CO 2 prediction is more sensitive to input than model selection. In this study, three novel machine learning algorithms of additive regression-random forest (AR-RF), Iterative Classifier Optimizer (ICO-AR-RF), and multi-scheme (MS-RF) were explored for carbon dioxide (CO 2 ) flux rate prediction from three agricultural fields. To build the dataset, 401 samples were collected from two fields in Prince Edward Island (PEI) and 122 samples from the New Brunswick (NB), Canada. In addition, soil moisture (SM), temperature (ST), and electrical conductivity (EC), alongside eight climatic variables including wind speed (WS), solar radiation (SR), relative humidity (RH), precipitation (PCP), air temperature (AT), dew point (DP), vapour pressure difference (VPD) and reference evapotranspiration (ET o ) were also collected. Greedy stepwise (GS) approach was implemented for feature selection. Finally, different qualitative (scatter plot, line graph, Taylor diagram, box plot, and Rug plot), and quantitative (uncertainty analysis, root mean square error (RMSE), percent of BIAS (PBIAS), Nash Sutcliff efficiency (NSE) and RMSE-observations standard deviation ratio (RSR)) techniques were used for model evaluation and comparison. Results of feature selection approaches revealed that DP, AT, SM, and ST are the four most effective variables at CO 2 prediction in PEI, while AT, RH, DP, and ST are the most effective in the NB study area. For optimum input scenario, the GS algorithm was applied, and results showed that a combination of DP, AT, ST, SM, and ET o was the best for the PEI study area, while for NB, all input variables should be involved. Our analysis, for prediction of CO 2 fluxes, confirmed that the ICO-AR-RF model performed the best at both PEI (RMSE=0.70, NSE=0.76, PBIAS=-5.11, RSR=0.48) and NB (RMSE=0.74, NSE=0.75, PBIAS=3.23, RSR=0.50), followed by MS-RF and AR-RF. Uncertainty analysis showed that CO 2 prediction is more sensitive to input scenario selection than models in both study areas. Results revealed that climatic variables are more effective in CO 2 prediction than soil characteristics and the developed hybrid model ICO-AR-RF can be a promising tool for decision-makers and beneficial for stakeholders.