ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume V-3-2020
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., V-3-2020, 469–476, 2020
https://doi.org/10.5194/isprs-annals-V-3-2020-469-2020
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., V-3-2020, 469–476, 2020
https://doi.org/10.5194/isprs-annals-V-3-2020-469-2020

  03 Aug 2020

03 Aug 2020

OPTIMAL DATES FOR DECIDUOUS TREE SPECIES MAPPING USING FULL YEARS SENTINEL-2 TIME SERIES IN SOUTH WEST FRANCE

N. Karasiak1, M. Fauvel2, J.-F. Dejoux2, C. Monteil1, and D. Sheeren1 N. Karasiak et al.
  • 1DYNAFOR, Université de Toulouse, INRAE, Castanet-Tolosan, France
  • 2CESBIO, Université de Toulouse, CNES/CNRS/INRAE/IRD/UPS, Toulouse, France

Keywords: sentinel-2, satellite image time series, tree species, forest, map, biodiversity, spatial autocorrelation, France

Abstract. The free to use Sentinel-2 (S2) sensors with 5-day revisit time at high spatial resolution in 10 spectral bands is a revolution in the remote sensing domain. Including 6 spectral bands in the near infrared, with 3 dedicated for the red-edge (where the vegetation significatively increases), these european satellites are very promising for mapping tree species distribution at a national scale. Here, we study the contribution of three one-year S2 Satellite Image Time Series (SITS) for mapping deciduous species distribution in the southwest of France. The annual cycle of vegetation (called phenology) can contribute to the identification of tree species. For some specific dates, species can have different phenological behaviours (senesence, flowering…). To train and validate the maps, we used the Support Vector Machine algorithm with a spatial cross-validation method. To train the algorithm with the same number of samples per species, we decided to undersample each class to the smallest class using a K-means clustering method. Moreover, a Sequential Feature Selection (SFS) has been implemented to detect the optimal dates per species. Our results are promising with high accuracy for Red oak andWillow (average score of the three one-year respectively F1 = 0.99, F1 = 0.94) based on the optimal dates. However, it appears that the performances when using the each full SITS are far below the optimal dates models (average ΔF1 = 0.32). We did not find, except for Willow and Red oak, that the optimal dates were the same for each year. Perspectives is to find an algorithm robust to temporal or spectral noise and to smooth the time series.