TY - JOUR
T1 - Estimating estuarine primary production using satellite data and machine learning
AU - Xu, Min
AU - Hu, Chuanmin
AU - Najjar, Raymond G.
AU - Herrmann, Maria
AU - Briceno, Henry
AU - Barnes, Brian B.
AU - Roger Johansson, J. O.
AU - English, David
N1 - Funding Information:
This work was supported by the U.S. NASA Ocean Biology and Biogeochemistry program (NNX17AH03G) and Water Resources program (80NSSC19K1200). We thank NASA GSFC for providing MODIS data and the City of Tampa Bay Study Group for providing the in situ water quality data used in the analysis. We also thank the anonymous reviewers who provided valuable comments and suggestions to improve the manuscript.
Publisher Copyright:
© 2022
PY - 2022/6
Y1 - 2022/6
N2 - Net primary production (NPP) plays an important role in estuarine carbon cycling, which has been increasingly impacted by human activities and global climate change. Spatiotemporal trends of NPP in the open ocean have been well studied using satellite data and standard primary production algorithms such as the Vertically Generalized Production Model (VGPM), but these algorithms are generally not suitable for estuarine and coastal waters. Previous remote sensing studies on estuarine NPP mainly focused on the tuning of the standard VGPM based on extensive local in situ data. Here we make a first attempt to use machine learning algorithms to estimate NPP in an estuarine environment from satellite measurements. Tampa Bay, the largest estuary in Florida (United States), has abundant in situ measurements of NPP, although the spatiotemporal variability of NPP within the bay remains unrevealed. Combining these data with concurrent MODIS/Aqua image data, we developed and evaluated seven machine learning algorithms (support vector regression, random forest, decision tree, bagging, adaptive boosting, gradient tree boosting, and neural network), and applied the one with the least root mean square error (RMSE) and highest correlation coefficient to establish a time-series NPP record for Tampa Bay from 2002 to 2020. Results of the best performing algorithm (Pearson's r = 0.82, RMSE = 151.0 mgC m−2 d−1) showed substantial improvement over the standard or tuned VGPM (Pearson's r = 0.51, RMSE = 364.9 mgC m−2 d−1), using the same independent variables for NPP ranging from 157.5 to 1368.8 mgC m−2 d−1. MODIS NPP shows temporal variations that are largely driven by temperature: lowest values in winter, highest values in summer, and an increasing trend from 2003 to 2020, highlighting the impact of global warming on estuarine NPP. The spatial distribution of MODIS NPP shows higher values in Hillsborough Bay, Middle Tampa Bay and Lower Tampa Bay, and relatively lower values in Old Tampa Bay, a pattern that likely reflects differences in river discharge. The long-term NPP product derived from machine learning algorithms and satellite data can complement existing field-based monitoring programs and help to understand estuarine responses to climate changes and human impacts, and design relevant mitigation strategies.
AB - Net primary production (NPP) plays an important role in estuarine carbon cycling, which has been increasingly impacted by human activities and global climate change. Spatiotemporal trends of NPP in the open ocean have been well studied using satellite data and standard primary production algorithms such as the Vertically Generalized Production Model (VGPM), but these algorithms are generally not suitable for estuarine and coastal waters. Previous remote sensing studies on estuarine NPP mainly focused on the tuning of the standard VGPM based on extensive local in situ data. Here we make a first attempt to use machine learning algorithms to estimate NPP in an estuarine environment from satellite measurements. Tampa Bay, the largest estuary in Florida (United States), has abundant in situ measurements of NPP, although the spatiotemporal variability of NPP within the bay remains unrevealed. Combining these data with concurrent MODIS/Aqua image data, we developed and evaluated seven machine learning algorithms (support vector regression, random forest, decision tree, bagging, adaptive boosting, gradient tree boosting, and neural network), and applied the one with the least root mean square error (RMSE) and highest correlation coefficient to establish a time-series NPP record for Tampa Bay from 2002 to 2020. Results of the best performing algorithm (Pearson's r = 0.82, RMSE = 151.0 mgC m−2 d−1) showed substantial improvement over the standard or tuned VGPM (Pearson's r = 0.51, RMSE = 364.9 mgC m−2 d−1), using the same independent variables for NPP ranging from 157.5 to 1368.8 mgC m−2 d−1. MODIS NPP shows temporal variations that are largely driven by temperature: lowest values in winter, highest values in summer, and an increasing trend from 2003 to 2020, highlighting the impact of global warming on estuarine NPP. The spatial distribution of MODIS NPP shows higher values in Hillsborough Bay, Middle Tampa Bay and Lower Tampa Bay, and relatively lower values in Old Tampa Bay, a pattern that likely reflects differences in river discharge. The long-term NPP product derived from machine learning algorithms and satellite data can complement existing field-based monitoring programs and help to understand estuarine responses to climate changes and human impacts, and design relevant mitigation strategies.
UR - http://www.scopus.com/inward/record.url?scp=85144542863&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85144542863&partnerID=8YFLogxK
U2 - 10.1016/j.jag.2022.102821
DO - 10.1016/j.jag.2022.102821
M3 - Article
AN - SCOPUS:85144542863
SN - 1569-8432
VL - 110
JO - International Journal of Applied Earth Observation and Geoinformation
JF - International Journal of Applied Earth Observation and Geoinformation
M1 - 102821
ER -