ESDS Program

Spatio-Temporal Machine Learning and Cloud Computing for Predicting Dynamics of Global Vegetation Structure from Active Satellite Sensors

Principal Investigator: Sassan Saatchi (NASA's Jet Propulsion Laboratory)

The Global Vegetation Structure (GVS) aim is to create machine learning models to integrate data from various remote sensing technologies to study global vegetation structure. The Lidar sensors on satellites or from airborne campaigns provide direct measurements of vegetation vertical profile but are limited in availability and coverage. Radar and optical sensors, on the other hand, offer indirect estimates of vegetation structure with excellent global coverage. By combining the strengths of both technologies using machine learning we can map vegetation structure and its changes over time.

Project Objectives

  • Assess lidar sources from different instruments (space-borne and airborne) and create inter-calibration methods to use them in conjunction, when required
  • Create training datasets of multi-sensor data (optical, radar) at multiple coherent spatial resolutions
  • Evaluate machine learning methods to create wall-to-wall maps of vegetation structure on a global scale

The work involves several essential steps to achieve the desired outcomes and create training datasets and high-quality maps of vegetation structure:

  • Data Collection for lidar: This entails collecting existing airborne lidar campaign data. Additionally, validation and comparison of space-borne and airborne products are necessary. For space-borne products, inter-calibration from different instruments may be required. Once these steps are completed, the data can be aggregated and processed to create a high-quality target dataset at various resolutions. The limitations of each sensor (sensor technology, measurement issues, time discrepancies) have to be taken into account, when required
  • Preprocessing of Dense Predictors: Different methodologies for preprocessing need to be developed and applied for each of the input datasets on a global scale. The most straightforward is the aggregation to different resolutions and alignment on compatible grids. The input datasets used include Landsat and Sentinel-2 bands for optical imagery, L-band Advanced Land Observing Satellite (ALOS) phased-array type L-band synthetic aperture radar (PALSAR) mosaic to provide wall-to-wall coverage for radar data
  • Testing and Comparison of Machine Learning Models: Utilizing the diverse input data sources, different machine learning models can be tested and compared to determine the most effective approach. Drawbacks and limitations of each method can be established, enabling a more comprehensive evaluation and improvements to guarantee a consistent multiscale approach and a correct evaluation of the costs entailed by each methodology
  • Intercomparison with Ground Truth Airborne Datasets: The derived products, along with existing datasets, can be compared against the ground truth airborne datasets to establish a benchmark and enable future developments

By following these steps, the study aims to optimize data collection and processing for lidar, explore various machine learning models, and establish benchmarks for accuracy assessment. This comprehensive approach will enable the development of robust and reliable vegetation structure analysis methods on a global scale.

Major Accomplishments

  • We developed a methodology to combine the information from two different lidar instruments, namely the Ice, Cloud and land Elevation Satellite-2 (ICESat-2) and the Global Ecosystem Dynamics Investigation (GEDI) missions, to be able to fill the observation gap in GEDI data over boreal areas
  • Multiple machine learning models for estimating vegetation structure from the optical and radar imagery have been tested and applied on a global scale with results on par with the current state-of-the-art

Publications and Presentations

Lee, J., Favrichon, S., Mauceri, S., et al. (2023). Addressing Underestimation in Global Forest Structure Mapping. ESS Open Archive. January 3, 2023. doi:10.22541/essoar.167276451.10705079/v1

Yan, Y., et al. (2022). Mapping Changes of Global Above Ground Biomass of Live Vegetation from 2001-2021 Using Satellite Observations and Deep-Learning Models. American Geophysical Union (AGU) Fall Meeting 2022.

Last Updated