Modelling effects of forest disturbance history on carbon balance: a deep learning approach using Landsat-time series.
Abstract
Forests play a crucial role in the global carbon (C) cycle, covering about 30% of the planet's terrestrial surface, accounting for 50% of plant productivity, and storing 45% of all terrestrial C. As such, forest disturbances affect the balance of terrestrial C dioxide (CO 2 ) exchange, with the potential of releasing large amounts of C into the atmosphere. Understanding and quantifying the effect of forest disturbance on terrestrial C metabolism is critical for improving forest C balance estimates and predictions. Here we combine remote sensing, climate, and eddy-covariance (EC) data to study forest land surface-atmosphere C fluxes at more than 180 sites globally. We aim to enhance understanding of C balance in forest ecosystems by capturing the ecological carry-over effect of disturbance historyon C fluxes. Our objectives are to (1) characterize forest disturbance history through the full temporal depth of the Landsat time series (LTS); and (2) to investigate lag and carry-over effects of forest dynamics and climate on ecosystem C fluxes using a data-driven recurrent neural network(RNN). The resulting data-driven model integrates carry-over effects of the system, using LTS, ecosystem productivity, and several abiotic factors. In this study, we show that our RNN algorithm is able to effectively calculate realistic seasonal, interannual, and across-site C flux variabilities based on EC, LTS, and climate data. In addition, our results demonstrate that a deep learning approach with embedded dynamic memory effects offorest dynamics is able to better capture lag and carry-over effects due to soil-vegetation feedback compared to a classic approach considering only the current condition of the ecosystem. Our study paves the way to produce accurate, high resolution carbon fluxes maps, providing morecomprehensive monitoring, mapping, and reporting of the carbon consequences of forest change globally.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2017
- Bibcode:
- 2017AGUFM.B31A1967B
- Keywords:
-
- 0428 Carbon cycling;
- BIOGEOSCIENCES;
- 0434 Data sets;
- BIOGEOSCIENCES;
- 0466 Modeling;
- BIOGEOSCIENCES;
- 0480 Remote sensing;
- BIOGEOSCIENCES