Seasonal Challenges for Rainforest Mapping with Sentinel-1 Time Series: A Deep Learning Approach

Abstract:
Tropical forests are extremely complex ecosystems, threatened by forest degradation and deforestation caused by human activity as well as by natural hazards. Thus, the availability of reliable and up-to-date data describing forest parameters is crucial in the effort to mitigate the disruption of these environments. A potential data source up to this task arises from spaceborne synthetic aperture radar (SAR) systems, whose imaging capabilities surpass those of optical sensors, in the sense that they are operational around-the-clock even with cloudy atmospheric conditions, which is typically the case of tropical rainforests. Sentinel-1 InSAR short time series have been successfully used as input to land cover classification problems with deep learning models on regional scale. In this paper, we propose to investigate the temporal information from Sentinel-1 repeat-pass interferometric SAR (InSAR) data, by considering time series with a temporal baseline smaller than a month, i.e., allowing for the generation of a forest map every 24 days. Moreover, we focus on possible ways for generalizing the proposed baseline approach to an operational framework for year-round forest mapping at a larger scale by investigating how seasonal variabilities affect the input data. Preliminary results in the Brazilian state of Rondonia indicate that employing additional information such as imaging acquisition dates might greatly improve the classification accuracy in forest areas with well-defined seasons.