Day 4

Detailed paper information

Back to list

Paper title Tropical Dry Forest Change Detection Using Sentinel Images and Deep Learning
Authors
  1. Adugna G. Mullisa Wageningen University & Research Speaker
  2. Diego Marcos Wageningen University & Research
  3. Bart Slagter Wageningen University
  4. Robert Masolele Wageningen University and Research
  5. Martin Herold GFZ German Center for Geosciences (Potsdam, Germany)
  6. Johannes Reiche Wageningen University
Form of presentation Poster
Topics
  • C1. AI and Data Analytics
    • C1.04 AI4EO applications for Land and Water
Abstract text Tropical Dry Forest Change Detection Using Sentinel Images and Deep Learning


Tropical dry forests (TDF) cover approximately 40% of the globally available tropical forest stock and play an essential role in controlling the interannual variability in the global carbon cycle, water cycle maintenance, reducing erosion and providing economic and societal benefits. Therefore, there is a strong need to persistently monitor changes in TDF to support sustainable land management and law enforcement activities to reduce illegal degradation. Satellite-based monitoring systems are the primary tools for providing information on newly deforested areas in vast and inaccessible forests. Recently, a temporally dense combination of optical and SAR images were used to counter the near constant cloud cover in tropical regions and increase early detection of deforestation events.


However, existing approaches and operational systems for satellite-based near real-time forest disturbance detection and monitoring such as the GLAD alerts (Hansen et al. 2016) and RADD alerts (Reiche et al. 2021) have mainly been used over tropical humid forests (THF) and their efficacy over TDF is largely undetermined because of the seasonal nature of TDF. Therefore, expanding the success of mapping capability from THF to TDF is of paramount importance. To this end, Combining optical and SAR datasets requires different methods for accurate inference as the observables are different due to the differences in image acquisition modality i.e. optical and SAR images observe different aspects of forest structures. In addition, utilizing optical and SAR images for TDF mapping requires robust seasonality mitigation to avoid false detections.


We will demonstrate a robust and accurate deep learning (DL) approach to map TDF changes from Sentinel-1 SAR and Sentinel-2 optical images. The designed DL approach utilizes a two-step weakly supervised learning framework. In the first step, it uses pixels where the Hansen annual forest change and GLAD alerts agree as initial reference of highly confident alerts. We then apply a hard positive mining strategy by searching for the earliest low confidence alerts at those same locations, which will be used to generate the labels to train our DL model. In the second step, the framework uses a Neural Network (NN) architecture with a self-attention mechanism to accurately infer TDF changes. This NN framework focuses on certain parts of the input sequences of images to allow for more flexible interactions between the different time steps in the image stack. The output from this framework will be compared with the output from standard recurrent neural networks such as the long-short term memory (LSTM) recurrent NN.



Hansen, M.C., Krylov, A., Tyukavina, A., Potapov, P.V., Turubanova, S., Zutta, B., Ifo, S., Margono, B., Stolle, F., Moore, R., 2016. Humid tropical forest disturbance alerts using Landsat data. Environ. Res. Lett. 11, 34008.
Reiche, J., Mullissa, A., Slagter, B., Gou, Y., Tsendbazar, N-E., Odongo-Braun, C., Vollrath, A., Weisse, M. J., Stolle, F., Pickens, A., Donchyts, G., Clinton, N., Gorelick, N., Herold, M. (2021) Forest disturbance alerts for the Congo Basin using Sentinel-1. Environmental Research Letters 16, 2, 024005. https://doi.org/10.1088/1748-9326/abd0a8.