The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLIII-B3-2021
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B3-2021, 559–566, 2021
https://doi.org/10.5194/isprs-archives-XLIII-B3-2021-559-2021
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B3-2021, 559–566, 2021
https://doi.org/10.5194/isprs-archives-XLIII-B3-2021-559-2021

  29 Jun 2021

29 Jun 2021

EXPLOITATION OF SPECTRAL AND TEMPORAL INFORMATION FOR MAPPING PLANT SPECIES IN A FORMER INDUSTRIAL SITE

R. Gimenez1,2, G. Lassalle3, R. Hédacq3, A. Elger2, D. Dubucq3, A. Credoz3, C. Jennet3, and S. Fabre1 R. Gimenez et al.
  • 1ONERA DOTA BP74025 2 av. Edouard Belin, FR-31055 Toulouse cedex 4, France
  • 2Laboratoire d’Ecologie Fonctionnelle et environnement, ENSAT - Avenue de l'Agrobiopole 31326 Castanet-Tolosan cedex, France
  • 3TOTAL S.E., CSTJF - avenue Larribau, 64000 Pau, France

Keywords: Hyperspectral, multispectral, multitemporal, vegetation, classification, anthropogenic impacts, machine learning

Abstract. Characterization and seasonal (periodic) monitoring of plant species distribution in the context of former industrial activity is crucial to assess long-term anthropogenic footprint on vegetated area. Species discrimination has shown promising results using both HyperSpectral (HS) and MultiSpectral (MS) images. Airborne HS instruments enable high spatial and spectral resolution imagery while time series of satellite MS images provide high temporal resolution and phenological information. This paper aims to compare supervised classification results obtained with non-parametric (Random Forest, RF, Support Vector Machine, SVM) and parametric methods (Regularized Logistic Regression, RLR) applied on both kinds of images acquired on an industrial brownfield. The studied site is a complex vegetated environment due to species diversity: 8 dominant species are retained. The performance obtained by preliminary feature selection based on principal component analysis and vegetation indices, to improve separability of spectral or temporal information according to species, is analysed. The best performance is obtained by RLR method applied on HS data without feature selection (global accuracy of 93 %). Feature selection is found to be a necessary step to perform classification with time series of MS images. Species that are difficult to distinguish from the HS image, namely Salix and Populus, are well separated using Sentinel-2 images (precision around 70%).