Day 4

Detailed paper information

Back to list

Paper title Sentinel-2 potential for Gross Primary Production estimation in forests by considering memory effects
Authors
  1. David Montero Loaiza Remote Sensing Center for Earth System Research (RSC4Earth), Leipzig University Speaker
  2. Miguel Mahecha Universität Leipzig
  3. Alexander Knohl University of Göttingen
  4. Sebastian Wieneke Universiät Leipzig
Form of presentation Poster
Topics
  • C1. AI and Data Analytics
    • C1.07 ML4Earth: Machine Learning for Earth Sciences
Abstract text Forests play a major role in the global carbon cycle and the mitigation of climate change effects. Gross Primary Production (GPP), the gross uptake of CO₂, is a key variable that needs to be accurately monitored to understand terrestrial carbon dynamics. Even though GPP can be derived from Eddy Covariance (EC) measurements at ecosystem scale (e.g., the FLUXNET network), the corresponding monitoring sites are sparse and unevenly distributed throughout the world. Data-driven techniques are among the most used methods to estimate GPP and its spatio-temporal fluctuations for locations where local measurements are unavailable. These methods entail developing an empirical model based on ground-truth GPP measurements and have been a primary tool for upscaling the GPP derived from EC measurements by using traditional Machine Learning methods with satellite imagery and meteorological data as inputs. Current data-driven carbon flux models utilize traditional models like Linear Regression, Random Forests, Support Vector Machines or Gaussian Processes, while Deep Learning approaches that leverage the temporal patterns of predictor variables are underutilised. Short- and long-term dependencies on previous ecosystem states are complex and should be addressed when modeling GPP. These temporally lagged dependencies of vegetation states, hereinafter memory effects, can be considered in traditional Machine Learning approaches, but must be encoded in hand-designed variables that lose their sequential structure. Here we show that the estimation of GPP in forests can be improved by considering memory effects using Sentinel-2 imagery and Long Short Term Memory (LSTM) architectures. We found that the accuracy of the model increased by considering the long-range correlations in time series of Sentinel-2 satellite imagery, outperforming single state models. Furthermore, the additional information contributed from Sentinel-2, such as its high spatial resolution (10-60 m), as well as the vegetation reflectance in the Red Edge bands (703-783 nm), boosted the accuracy of the model. Our results demonstrate that long-term correlations are a key factor for GPP estimation in forests. Moreover, the Red Edge reflectance enhances the sensitivity of the model to photosynthetic activity and the high spatial resolution of the imagery allows to account for local spatial patterns. These results imply that novel data-driven models should account for long-term correlations in remote sensing data. Additionally, the information provided by Sentinel-2 imagery demonstrated to increase the accuracy of the model and further investigation should be carried out. For example, local spatial patterns (e.g., tree mortality or deforestation in certain spots in the image) can be further exploited by Deep Learning methods such as Convolutional Neural Networks (CNN).