Towards the Monitoring of the Amazon Rainforest with TanDEM-X and Deep Learning Strategies

Kurzfassung

In the actual study for forest mapping over the Amazon rainforest, a U-Net has been improved by including as input feature relevant information on the acquisition geometry, such as height of ambiguity (related to the perpendicular baseline) and the local incidence angle. Moreover, given the special environment presented by the Amazon region, with the presence of many river beds, the U-Net has also been extended for multi-layer semantic segmentation, providing three classes: forest, non-forest, and water. The U-Net has been trained from scratch to avoid any type of transfer learning from previous works, by implementing an ad-hoc strategy with allows the model to generalize well on all different acquisition geometries. Mainly images acquired in 2011 have been used for the training, to minimize the temporal distance to the used independent reference, a forest map based on Landsat data from 2010. TanDEM-X images acquired in 2012 have been considered too, to account for the high variability in the interferometric acquisition geometries. In total, 455 images have been used for training and 320 for testing, covering three ranges of incidence angles, as well as heights of ambiguity between 20 and 120 m. Moreover, the considered images present a forest content, according to the reference data, between 30% and 70% to account for a balanced class training. Extra images have been considered to train the U-Net on the detection of water. 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 for the TanDEM-X forest/non-forest map, with a f-score increase of 0.13. The improvement in the forest classification makes it possible 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 waterbodies, as done for the TanDEM-X forest/non-forest map. 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 Amazonas Rainforest over the years and to follow deforestation patterns. By increasing the number of TanDEM-X acquisitions over the Amazonas and applying the trained CNN it would be possible to perform a near real-time forest monitoring over selected hot-spot areas.