Project highlights

  • Satellite imagery makes it possible to monitor vegetation biophysical parameters such as biomass carbon and forest dynamics from space automatically using artificial intelligence;
  • Google’s TensorFlow AI software will be applied to Copernicus Sentinel-1, Sentinel-2, Planet and JAXA’s ALOS-2 PALSAR-2 satellite imagery;
  • Expected outcomes include the automated detection and labelling of different types of forest dynamics and the analysis of aboveground biomass carbon trends.

Overview

Amazon forests hold the largest pools of forest biomass, but this estimate remains poorly quantified. Remote sensing methods provide the necessary tools to quantify forest biomass at large scales, although the local forest inventories data availability usually constrains them. There are notable differences among current biomass maps (Rodriguez-Veiga, et al., 2017) because of forest inventories’ availability, leading to high carbon emissions uncertainties when referring to specific biomass maps (Figure 1). Therefore, we urge a new remote sensing approach to estimate the aboveground biomass of tropical forests without relying heavily on forest inventory data.

Recent advances in high-performance computing and ‘cloud-data’ are paralleled with advanced artificial intelligence algorithms and significant investment in new satellite missions. Machine learning (Le Cun et al. 2015) have previously been applied to hyperspectral image classification (Hu et al. 2015), CORINE land cover mapping from Sentinel-1 SAR images (Balzter et al. 2015), and forest biomass mapping using a combination or SAR and optical images (Rodriguez-Veiga et al, 2016). Machine learning enables automatic detection of forest changes of satellite images. The paradigm of looking for spatial and temporal patterns instead of the historic focus on spectral information in satellite imagery allows for identification of different types of forest dynamics (e.g. disturbance and succession). Artificial intelligence can also be used to accurately estimate from space, forest biophysical parameters that are difficult to measure in forest inventory data (Rodriguez-Veiga et al, 2017) (Figure 1).

This interdisciplinary studentship aims to explore the use of machine learning to quantify changes of aboveground biomass carbon in dense time-series satellite data of several Brazilian forest sites. Time-series stacks of multispectral optical and synthetic Aperture Radar (SAR) sensors will be input into the AI. The AI will be trained based on measurements collected from in-situ forest inventories and visual interpretation of very high resolution images.

Research questions:

  1. What are the carbon stocks and fluxes of aboveground biomass of the Amazon forests?
  2. How accurately can artificial intelligence be trained to quantify forest dynamics of the Amazon based on satellite time-series information?
  3. How accurately can the associated aboveground biomass loss or gain be estimated?

Host

University of Leicester

Theme

  • Climate and Environmental Sustainability
  • Organisms and Ecosystems

Supervisors

Project investigator

  • Fernando Espirito-Santo, School of Geography, Geology and the Environment, University of Leicester

 

Co-investigators

  • Prof. Heiko Balzter, National Centre for Earth Observation (NCEO), University of Leicester
  • Dr. Pedro Rodriguez-Veiga, National Centre for Earth Observation (NCEO), University of Leicester

How to apply

Methodology

Methods will be drawn from mathematical modelling, artificial intelligence and earth observation over a number of sites of the Brazilian Amazon region. The TensorFlow AI (Abadi et al. 2016) will be implemented on the high-performance computing facility SPECTRE-2 at the University of Leicester and linked with an existing 10-m resolution Sentinel-2 image processing chain developed in Python (https://github.com/clcr/pyeo). The project will also include others images from Planet, Sentinel-1 and ALOS-2 PALSAR-2 satellites. Training and validation data for the AI will be available from the Global Ecosystem Dynamics Investigation (GEDI) Lidar footprints (https://gedi.umd.edu/), in-situ forest inventories, and interpretation of high resolution imagery. Once trained up, the AI will identify forest dynamics types and aboveground biomass loss and gain over time. Experiments with different spatial filters and configurations of TensorFlow will be undertaken to optimise the detection of particularly difficult forest dynamics (e.g. gold mining, regrowth, under canopy crops), and estimation of high forest biomass density levels.

Training and skills

The student will be trained in Sentinel data processing on the HPC facility SPECTRE-2 at University of Leicester. The student will take the new MSc module GY7709 (Satellite Data Analysis in Python), available since 2019-2020, and any other modules deemed suitable, dependent on the background of the student. Complementary, individual training in using AI, especially TensorFlow, will be available from the Department of Mathematics. Further training will take place ‘on-the-job’ as part of the research team.

Partners and collaboration

  • Potential partners include (but not limited to) the Brazilian Space Agency (INPE)
  • University of São Paulo and UNICAMP;
  • Planet – satellite assembly, operations and data analytics company
  • Google – London office has expressed interest in projects around AI and satellite data

Further details

Please visit the University of Leicester website for application guidance:

https://le.ac.uk/study/research-degrees/funded-opportunities/centa-phd-studentships

Possible timeline

Year 1

Literature review, refinement of research questions and work plan, liaison and consultation with project partners, installation of AI on HPC, training in satellite data processing in Python

Year 2

Preparation of training and validation database (GEDI, in-situ, and HR imagery). Training the AI, completing test runs with optical data (Sentinel-2 and Planet) and evaluating outcomes, iteratively refining data flow and accuracy, running AI ‘experiments’

Year 3

Implementing SAR data (ALOS-2 PALSAR-2 and Sentinel-1) AI runs, evaluating and comparing results, submitting 2 papers for publication

Further reading

Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M., Ghemawat, S., Irving, G., Isard, M. and Kudlur, M. (2016): Tensorflow: a system for large-scale machine learning. OSDI 16, 265-283, https://www.usenix.org/system/files/conference/osdi16/osdi16-abadi.pdf

Balzter, H., Cole, B., Thiel, C. and Schmullius, C. (2015): Mapping CORINE Land Cover from Sentinel-1 SAR and SRTM Digital Elevation Model using Random Forests. Remote Sensing 7, 14876-14898. https://doi.org/10.3390/rs71114876

Cole, B., Smith, G. and Balzter, H. (2018): Acceleration and fragmentation of CORINE land cover changes in the United Kingdom from 2006-2012 detected by Copernicus IMAGE2012 satellite data. International Journal of Applied Earth Observation and Geoinformation 73, 107–122. https://doi.org/10.1016/j.jag.2018.06.003

Comber, A., Balzter, H., Cole, B., Fisher, P., Johnson, S. and Ogutu, B. (2016): Methods for quantifying regional differences in land cover change. Special Issue on Validation and Inter-Comparison of Land Cover and Land Use Data, Remote Sensing 8, 176-195. https://doi.org/10.3390/rs8030176

Hu, W., Huang, Y., Wei, L., Zhang, F. and Li, H. (2015): Deep convolutional neural networks for hyperspectral image classification. Journal of Sensors 2015, 258619, https://doi.org/10.1155/2015/258619

Le Cun, Y., Bengio, Y. and Hinton, G. (2015): Deep learning. Nature, 521(7553), 436. https://doi.org/10.1038/nature14539

Le Toan, T., S. Quegan, M. W. J. Davidson, H. Balzter, P. Paillou, K. Papathanassiou, S. Plummer, F. Rocca, S. Saatchi, H. Shugart and L. Ulander (2011): The BIOMASS mission: Mapping global forest biomass to better understand the terrestrial carbon cycle. Remote Sensing of Environment 115(11): 2850-2860. https://doi.org/10.1016/j.rse.2011.03.020

Rodríguez-Veiga, P., Saatchi, S., Tansey, K., and Balzter, H. (2016): Magnitude, spatial distribution and uncertainty of forest biomass stocks in Mexico. Remote Sensing of Environment. 2016;183:265-81. http://dx.doi.org/10.1016/j.rse.2016.06.004

Rodríguez-Veiga, P., Wheeler, J., Louis, V., Tansey, K., and Balzter, H. (2017): Quantifying Forest Biomass Carbon Stocks From Space. Current Forestry Reports 3: 1. https://doi.org/10.1007/s40725-017-0052-5

Rodríguez-Veiga, P., S. Quegan, J. Carreiras, H. J. Persson, J. E. S. Fransson, A. Hoscilo, D. Ziółkowski, K. Stereńczak, S. Lohberger, M. Stängel, A. Berninger, F. Siegert, V. Avitabile, M. Herold, S. Mermoz, A. Bouvet, T. Le Toan, N. Carvalhais, M. Santoro, O. Cartus, Y. Rauste, R. Mathieu, G. P. Asner, C. Thiel, C. Pathe, C. Schmullius, F. M. Seifert, K. Tansey and H. Balzter (2019): Forest biomass retrieval approaches from earth observation in different biomes. International Journal of Applied Earth Observation and Geoinformation 77: 53-68. https://doi.org/10.1016/j.jag.2018.12.008

 

COVID-19

Current COVID-19 pandemic might affect project delivery due to the following:

  • Travel disruptions.
    • This problem can be mitigated by video-conference software to hold meetings and training events.
    • The project is not reliant on fieldwork as it uses primarily digital data.
  • Slow internet causing disruptions in video-conference meetings:
    • This may affect partners and can be mitigated by holding virtual meetings at times when internet traffic is at a minimum
  • Prevention of face-to-face meetings due to national/local lockdowns
    • The project can be run entirely remotely if needed. Regular check-ins online are scheduled. Project design to remain entirely digital, no physical experiments.
  • PhD training cannot take place in person
    • Teaching and learning is now taking place following the blended learning approach under IGNITE, which can be switched to completely online teaching under conditions of restrictions and includes asynchronous and synchronous delivery.