ObjectivesThe hyperspectral images (HI) is at the moment still too poorly considered; nevertheless its specificities make a weighty auxiliary for the monitoring of the elements of the urban area. The HYEP project has for objective to propose a panel of methods and processes designed for hyperspectral imaging. We take into account other existing sensors in order to compare the performances. If the IH is complementary to the sensors HRS and VHRS due to its better radiometric richness, it allows to identify and to characterize the natural or anthropogenic elements in a complementary perspective. To this end methods for the extraction of information had to be adapted, created even. The methodological part of the project articulated at the same time in the solidification of the current approaches and the design of new methods. Results have been presented along the project duration to scientific community and local authorities. One of the milestones of the works was the comparison of the results to various spatial resolutions to specify the contribution of such a hyperspectral sensor with regard to those existing or to come. Data and data processing Methods: existing or adapted The methods were chosen among all the existing methods by adapting them to the signal, among spectral ranges and to the characteristics of urban areas. Since data arose from airborne platforms, the first developments were realized to counter the effects of the atmosphere on the IH (atmospheric correction - 3 tested methods) and a database of spectral signature for diverse elements of land use in town (roofs, roads, the vegetation etc.) was established. It allowed encircling better the spectral values of materials. It was set up based simultaneously on the literature, in situ and laboratory measurements. Its contributions in various classification processing were tested. Methods for information, extraction, pansharpening or classification purposes were used for various spatial and spectral resolutions to highlight its interest towards other sensors and also its benefits for a spatial mission. Classification and unmixing methods have been adapted and spatial pattern indicators for urban areas defined.Outcomes- 3 atmospheric correction methods have been tested; it leads to a specific code development by ONERA.- Methods adaptation : pansharpening and unmixing- Transfer: a complete design of the study has been transfered to Kaunas (Lituania) teams- Algorithms: Depository on http://openremotesensing.net/- One of the major results is the extraction and the identification of photovoltaic panel- CNRS Summer school 2017Scientific productionThe team has presented at ISPRS Geospatial Week 2015, GeoHyper (2015), Jurse 2017, IGARSS 2016-2017, SFPT or workshops TEMU, AFIGEO and to GdR Session (MaDics and ISIS) or within the framework of the Hypxim mission. The team organized special conference sessions at the national level, SFPT hyperspectral (2016) and international level IGARSS 2018 and WHISPERS 2018. A thematic CNRS summer school (2017 August 28 - September 1st - 25 participants) has been set up.The project gave rise to 10 publications Rang A and 38 communications, 4 chapters and a special issue for the RemoteSensing journal.

Hyperspectral data were obtained during an acquisition campaign led on Toulouse (France) urban area on July 2015 using Hyspex instrument which provides 408 spectral bands spread over 0.4 – 2.5 μ. Flight altitude lead to 2 m spatial resolution images. Supervised SVN classification results for 600 urban trees according to a 3 level nomenclature: leaf type (5 classes), family (12 and 19 classes) and species (14 and 27 classes). The number of classes differ for the two latter as they depend on the minimum number of individuals considered (4 and 10 individuals per class respectively). Trees positions have been acquired using differential GPS and are given with centimetric to decimetric precision. A randomly selected subset of these trees has been used to train machine SVM and Random Forest classification algorithms. Those algorithms were applied to hyperspectral images using a number of classes for family (12 and 19 classes) and species (14 and 27 classes) levels defined according to the minimum number of individuals considered during training/validation process (4 and 10 individuals per class, respectively). Global classification precision for several training subsets is given by Brabant et al, 2019 (https://www.mdpi.com/470202) in terms of averaged overall accuracy (AOA) and averaged kappa index of agreement (AKIA).