Day 4

Detailed paper information

Back to list

  1. José-Luis Bueso-Bello German Aerospace Center (DLR) e.V. Speaker
  2. Daniel Carcereri DLR - Deutsches Zentrum für Luft- und Raumfahrt e.V.
  3. Paola Rizzoli DLR - Deutsches Zentrum für Luft- und Raumfahrt e.V.
Form of presentation Poster
  • C1. AI and Data Analytics
    • C1.04 AI4EO applications for Land and Water
Abstract text The Amazon rainforest is the largest moist broadleaf tropical forest on the planet and plays a key role for regulating environmental processes on the Earth. It is a crucial element in the carbon and water cycles and acting as climate regulator, e.g. by absorbing CO2 and producing about 20% of the Earth's oxygen and this way counteracting global warming. The monitoring of changes in such forested areas, as well as understanding the water dynamics in such a unique biome is of key importance for our planet. Synthetic Aperture Radar (SAR) systems, thanks to its capability to see through the clouds, are an attractive alternative to optical sensors for remote sensing over such areas, which are covered by clouds for most of the year.
From TanDEM-X acquisitions it is possible to derive amplitude as well as bistatic coherence images. By exploiting the interferometric coherence, and specifically the volume correlation factor, it is possible to distinguish forested areas from non-vegetated ones, as demonstrated for the generation of the global TanDEM-X Forest/Non-Forest Map, that was based on a supervised clustering algorithm [1]. The interferometric coherence was also the main input for global water mapping using a watershed segmentation algorithm, as shown in the production of the TanDEM-X Water Body Layer [2]. On both global products, provided at a resolution of 50 m x 50 m, it was necessary to mosaic overlapping acquisitions to reach a good final accuracy.
Deep learning methods, concretely the U-Net presented in [3], showed promising results to accurately distinguish forested areas on a limited set of single TanDEM-X full-resolution images at 12 m x 12 m. In the actual study for forest and water monitoring over the Amazon rainforest, this U-Net architecture has been used as base to extend the capabilities of deep learning methods to work with TanDEM-X images acquired with a larger variety of acquisition geometries and to provide large scale maps including forest and water detection. The height of ambiguity (related to the perpendicular baseline) and the local incidence angle have been included in the input features set as the main descriptor of the bistatic acquisition geometry. The U-Net has been trained from scratch to avoid any type of transfer learning from previous works, by implementing an ad-hoc strategy which allows the model to generalize well on all different acquisition geometries. Mainly images acquired in 2011 and 2012, representing the high variability in the interferometric acquisition geometries, have been used for the training, in order to minimize the temporal distance to the used independent reference, a forest map based on Landsat data from 2010. The selected images for training and validation of the U-Net, as well as the selected images for testing, cover the three ranges of imaging incidence angles as in [1], as well as heights of ambiguity between 20 m and 150 m. Special attention was paid to balance the three different classes, forest, non-forest and water, in each one of the combined ranges of imaging incidence angle and height of ambiguity.
By applying the proposed method on single TanDEM-X images, we achieved a significant performance improvement in the test images with respect to the clustering approach developed in [1], with a f-score increase of 0.13 for the forest class. An improvement of the classification of forest with the CNN is overall observable, but especially noticeable over dense forested areas (percentage of forest samples > 70%). Also, the classification approach with deep learning methods can be extended to images acquired with a height of ambiguity > 100 m, which was a limitation of the clustering approach shown in [1]. Indeed, with the clustering approach images acquired with high height of ambiguity values resulted in an ambiguous forest classification, due to the smaller perpendicular baselines between the satellites, which reduce the volume decorrelation.
Such improvements make it possible to extend the number of useful TanDEM-X images and allows us to skip the weighted mosaicking of overlapping images used in the clustering approach for achieving a good final accuracy at large scale. Moreover, no external references are necessary either to filter out water bodies, as for the forest/non-forest map in [1]. In this way, we were able to generate three time-tagged mosaics over the Amazon rainforest utilizing the nominal TanDEM-X acquisitions between 2011 and 2017, just by averaging the single image maps classified by the ad-hoc trained CNN. These mosaics can be exploited to monitor the changes over the Amazon rainforest over the years and to follow deforestation patterns and changes in river bed extensions. By increasing the number of TanDEM-X acquisitions over the Amazonas and by applying the trained CNN it will be possible to perform a near real-time forest monitoring over selected hot-spot areas and to easily extend such a classification approach to other tropical forest areas.

[1] M. Martone, P. Rizzoli, C. Wecklich, C. Gonzalez, J.-L. Bueso-Bello, P. Valdo, D. Schulze, M. Zink, G. Krieger, and A. Moreira, “The Global Forest/Non-Forest Map from TanDEM-X Interferometric SAR Data”, Remote Sensing of Environment, vol. 205, pp. 352–373, Feb. 2018.
[2] J.L. Bueso-Bello, F. Sica, P. Valdo, A. Pulella, P. Posovszky, C. González, M. Martone, P. Rizzoli. “The TanDEM-X Global Water Body Layer”, 13th European Conference on Synthetic Aperture Radar, EUSAR, 2021.
[3] A. Mazza, F. Sica, P. Rizzoli, and G. Scarpa, “TanDEM-X forest mapping using convolutional neural networks”, Remote Sensing MDPI, vol. 11, 12 2019.