B066-0022
Long-Short Term Memory Neural Network for gap-filling N2O field measurements
Friday, 11 December 2020
Poster
Christopher Dorich1, Daqi Wang1, Cameron Key2, Massimiliano De Antoni Migliorati3,4, Peter Grace3 and Richard T Conant1,5, (1)Colorado State University, Natural Resource Ecology Lab, Fort Collins, CO, United States, (2)Colorado State University, Fort Collins, United States, (3)Queensland University of Technology, Brisbane, Australia, (4)Queensland Government, Department of Environment and Science, Queensland, Australia, (5)Colorado State University, Ecosystem Science and Sustainability, Fort Collins, CO, United States
Abstract:
Nitrous oxide emissions are highly variable – changing based on site conditions, management practices, soil chemistry and conditions. Sampling of N
2O can also be time consuming and expensive, resulting in gaps within field data. Determining sampling strategies that do not bias cumulative emissions (e.g., peak chasing) and methods for gap-filling or estimating emissions between observations is thus necessary for improving emission factors, calculating cumulative emissions, and determining mitigation strategies. Advanced gap-filling methods (GAMs, ARIMA, Bayesian methods, Random Forest, and Neural Networks) have started to be explored in the N
2O community for their ability to estimate N
2O emissions or gap-fill field measurements. While these methods have shown promise, they have also generally been limited to testing on individual sites.
Using the Global N2O Database we assembled a group of sites (6) that sampled for more than 200 days over a yearlong period and contained accompanying covariate data (air temperature, precipitation, soil moisture and temperature, soil inorganic N). This high sampling frequency allowed for creation of artificial gaps within the data, providing for model testing where validation statistics could be calculated based on the testing data (observed N2O emissions) and model estimate. This assemblage of sites (with 24 Treatments) provides a larger data set, necessary for Neural Network (NN) methods, than has previously been accessible. Using these datasets we calibrated a Long-Short Term Memory (LSTM) NN that split the data into separate calibration, validation, and hold out data sets, to ensure the model would not over fit a data set and would be more generalizable. The model was built following a gap-filling procedure that iterated over various model architectures (NN learning rate, batch size, etc), testing scenarios (observed N2O provided every 3, 5, 7, 14 days), and covariate data. Here we present the results of the NN and discuss limitations and next steps of the model.
