Day 4

Detailed paper information

Back to list

Paper title A CNN-based approach for forest parameter regression by fusion of Sentinel-2 and TanDEM-X data
Authors
  1. Daniel Carcereri DLR - Deutsches Zentrum für Luft- und Raumfahrt e.V. Speaker
  2. Paola Rizzoli DLR - Deutsches Zentrum für Luft- und Raumfahrt e.V.
  3. Dino Ienco INRAE - UMR TETIS
  4. José-Luis Bueso-Bello German Aerospace Center (DLR) e.V.
  5. Carolina González DLR - Deutsches Zentrum für Luft- und Raumfahrt
  6. Lorenzo Bruzzone University of Trento
Form of presentation Poster
Topics
  • A3. Biosphere
    • A3.06 Biomass monitoring
Abstract text Forests cover an estimated 31% of the Earth's global surface and therefore constitute a significant part of the biosphere. They fundamentally impact the carbon cycle as vegetation is able to absorb carbon atoms from the atmosphere and store them, by building up new biomass during their natural growth process.
Forests also majorly affect the local water-cycle, as the transpiration process redistributes ground-water into the atmosphere, impacting air temperature and weather in the process.
Forests are also critical for biodiversity preservation, an estimated 80% of all known terrestrial flora and fauna lives in them. Similarly, about 880 million people collect and produce fuel from wood while 90% of people living in extreme poverty have their livelihoods depending on forests.
To accurately estimate forest tree parameters such as canopy height (CHM) and above ground biomass (AGB), it is common practice to measure them manually on-site. This process can be both invasive, when individual trees are cut down to precisely assess their properties, or non-invasive, when a less intrusive approach is preferred over absolute accuracy.
The process is very expensive and time consuming, especially in remote areas. Therefore, in-situ measurement campaigns are feasible only for small surveys.
Airborne LiDAR systems also remain impractical and expensive when both large scale and low revisit time measurements are required, while spaceborne ones do not allow yet for the retrieval of wall-to-wall measurements.
As a consequence, spaceborne imaging systems for earth observation (EO) have gained wide interest in the last decades, as a large list of sensors and techniques is available delivering remote-sensing data at very large scales and low revisit times. Since this does not directly quantify forest parameters, it is necessary to model the relationship between the acquired data and the on-ground forest parameters.
Allometric equations are commonly used to indirectly relate forest parameters with RS data, but they require parameters to be tuned to the specific forest types and geographic locations to achieve good performance.
More sophisticated, physics-based modelling approaches have also been studied for the regression of forest parameters.
These tend to achieve high accuracy in their estimates, while retaining great spatial resolution.
To obtain these results, large amounts of data, auxiliary information or ground reference samples are required to invert the models.
With the recent advancements in machine learning and computer vision techniques, and the availability of large dataset collections from EO sensors, new approaches to forest parameter regression are starting to be explored.
Deep learning architectures have already found great success for classification tasks, as they analyze the spatial context information to generate higher level abstractions, producing features which typically possess a larger descriptive and discriminative content than both the input imagery and hand-crafted features.
On the other hand, comparatively little work still exists regarding the regression of physical and biophysical parameters from RS data, presumably due to the limited availability of large quantities of reference-data required for supervised training.
Aiming at providing large-scale, frequently updated CHM and AGB forest parameter metrics, our research effort focuses on overcoming the aforementioned limitations by proposing a multi-modal CNN-based regression framework, requiring only a single set of either single- or multi-source satellite imagery as input.
This multi-sensor approach represents a flexible solution for the continuous monitoring of forests when one or more input data sources are unavailable, and to otherwise achieve the best possible performance. In particular, we focus on combining high resolution Sentinel-2 optical imagery with TanDEM-X-derived interferometric SAR (InSAR) products, as they both provide fundamentally complementary information, and have been demonstrated to correlate well with forest parameters. The proposed data-driven multi-sensor approach consists in a deep multi-branch CNN architecture, where each of the modalities is associated to a separate feature extraction (encoder).
The spatial context extracted from these branches is then fused to supply a rich set of input features to a shared regression branch. We use a so-called cross-fusion approach to do this, which consists in a dedicated convolutional architecture that fuses different modalities through a set of convolutions and concatenations.
To assess the capabilities of the multi-branch architecture to fuse Sentinel-2 and TanDEM-X data and the regression performance of our framework, four tropical regions in Gabon, Africa have been considered. These correspond to reference data that has been acquired in the context of the 2016 AfriSAR campaign and consist of AGB maps, which have been derived at a ground sampling distance of 50m from airborne LiDAR measurements by fitting allometric equations on specifically acquired field-plot measurements.
We expanded the analysis period from mid 2015 to early 2017, since in 2016 only one Sentinel-2 satellite was available, which, combined with the extended cloud coverage over tropical regions, meant that only a small amount of imagery would have been available. We assumed that the changes in biomass are negligible within this time frame, as mainly tropical primary forest is considered.
During the learning phase, the network was trained on 32x32 pixel patches, using the mean square error (MSE) of the prediction as loss function for the backpropagation step. A validation set was used to select the best performing network across 10 training iterations. Finally, a separate test set was used to provide unbiased accuracy assessments.
Preliminary results in Gabon using Sentinel-2 optical and TanDEM-X interferometric SAR products are promising, showing agreement with the underlying assumptions and expectations. The root mean square error (RMSE) obtained on the test set is equal to 70.2 Mg/ha with a coefficient of determination R²=0.73, which is in line with the state-of-the-art methods.
We expect further optimization of the network and a more representative data set for training to further improve the estimation accuracy, setting the ground floor for the establishment of an effective tool for monitoring forest resources.