Burned Area Mapping Using SAR and Multispectral Data Integration via Generative Adversarial Networks

Multispectral data have been successfully exploited in the literature for burned area detection. However, the occurrence of suitable weather and illumination conditions limits their availability, especially at high latitudes and tropical environments. The usage of synthetic aperture radar (SAR) data, due to their all-weather and all-time characteristics, can potentially be a breakthrough, but it is still underexploited due to the ambiguity of the radar reflectivity of burned vegetation. The recent availability of generative networks is providing a new tool for combining such complementary information. This paper introduces a new framework for burned area mapping based on image domain translation. The main novelties against the related literature are the i) exploitation of SAR phase information as an input for the translation and the pointing to multispectral indices as a translation target. In particular, a bi-temporal coherent change detection-oriented SAR product is used as input for generative networks targeting the post-event normalized burned ratio index. Then, generated data are combined with native optical pre-event ones to produce the difference image typically used for thresholding segmentation. Experimental results obtained using a dataset extracted from the Copernicus Emergency Management Service database show that the proposed methodology outperforms the literature reaching a burned area mapping accuracy of above 90% with negligible false alarm, thus introducing a new robust solution for the integration of SAR and multispectral data.