Machine Learning and Pattern Analysis Identify Distinctive Influences from Long-term Weekly Net Ecosystem Exchange at Four Deciduous Woodland Locations

A methodology integrating correlation, regression (MLR), machine learning (ML), and pattern analysis of long-term weekly net ecosystem exchange (NEE) datasets are applied to four deciduous broadleaf forest (DBF) sites forming part of the AmeriFlux (FLUXNET2015) database. Such analysis effectively characterizes and distinguishes those DBF sites for which long-term NEE patterns can be accurately predicted using the recorded environmental variables, from those sites cannot be so delineated. Comparisons of twelve NEE prediction models (5 MLR; 7 ML), using multi-fold cross-validation analysis, reveal that support vector regression generates the most accurate and reliable predictions for each site considered, based on fits involving between 16 and 24 available environmental variables. SVR can accurately predict NEE for datasets for DBF sites US-MMS and US-MOz, but fail to reliably do so for sites CACbo and MX-Tes. For the latter two sites the predicted versus recorded NEE weekly data follow a Y ≠ X pattern and are characterized by rapid fluctuations between low and high NEE values across leaf-on seasonal periods. Variable influences on NEE, determined by their importance to MLR and ML model solutions, identify distinctive sets of the most and least influential variables for each site studied. Such information is valuable for monitoring and modelling the likely impacts of changing climate on the ability of these sites to serve as long-term carbon sinks. The periodically oscillating NEE weekly patterns distinguished for sites CA-Cbo and MX-Tes are not readily explained in terms of the currently recorded environmental variables. More detailed analysis of the biological processes at work in the forest understory and soil at these sites are recommended to determine additional suitable variables to measure that might better explain such fluctuations