Mapeamento automática de plantas exóticas invasoras a nível nacional recorrendo a uma abordagem de deteção remota de multi-resolução

MAP.INVADER project will address an important challenge which is not currently addressed in the literature which is the mapping of an IAP (i.e. A.dealbata) over national extents. Several difficulties contribute to this such as: 1) current image recognition models need a vast amount of quality training data deemed representative of the target class; 2) spectral similarities between IAP in general and also A.dealbata in particular, with other vegetation land cover classes; 3) national mapping of land cover needs to deal with the different landscape characteristics at a national level; 4) manual labelling of synoptic satellite imagery is still a challenge which is often based on extensive field work campaigns; 5) there is no economic incentive in IAP mapping over national extents. MAP.INVADER main objective is to propose an A.dealbata national mapping framework to generate yearly national maps over the period of the project for continental Portugal. The aim will be to assess the use of the UAV and VHR imagery not to directly map occurrences of A.dealbata but instead use these data as input for the generation of training for a national Sentinel-2 image segmentation framework. Hence, in a first step several areas will be selected as representative of the Portuguese landscape. These will then be surveyed with either UAV or VHR imagery. Image segmentation models will be trained to generate A.dealbata maps over the selected reference sites. These reference sites will then be used to directly annotate the decameter resolution Sentinel-2 images, deriving the needed training data for the national mapping of A.dealbata. To achieve this, several research problems will be addressed: 1) optimal reference site selection, as such reference sites need to be maintained to a minimum, reducing associated costs (UAV survey or buying VHR imagery); while being representative of the Portuguese landscape; 2) reduce the manual annotation effort for the UAV and VHR imagery concerning the reference sites selected in point 1, by taking advantage of weakly supervised deep learning algorithms to train UAV and VHR image segmentation approaches; 3) overall national mapping framework of A.dealbata by having the reference sites selected in point 1, classified with the models coming from point 2, which will used to directly label Sentinel-2 images for a posterior training of an image segmentation model capable of performing over continental Portugal. The operationalization of such research is also central with MAP.INVADER. This is aimed at making the developed research items indicated above be integrated into a framework which is able not only to address these research problems but at the same time be useful for several Portuguese institutions. To this regard, MAP.INVADER is counting with the support of several Portuguese institutes: Portuguese Institute for Nature and Conservation (ICNF), Portuguese Directorate-General for the Territory (DGT) and Intermunicipal Community of Coimbra’s Region (CIM-RC); to guide the operationalization of the overall framework to match end-users needs.