High Dimensional Kullback-Leibler Divergence for grassland object-oriented classification from high resolution satellite image time series

Mailys Lopes 1 Mathieu Fauvel 1 Stéphane Girard 2 David Sheeren 1 Marc Lang 1
2 MISTIS - Modelling and Inference of Complex and Structured Stochastic Systems
Inria Grenoble - Rhône-Alpes, LJK - Laboratoire Jean Kuntzmann, INPG - Institut National Polytechnique de Grenoble
Abstract : In rural landscapes, grasslands are a major source of biodiversity. They provide many ecosystem services. In particular, they are the main livestock feeding resource, through grazing and forrage production. Semi-natural grasslands host a rich species composition that is impacted by the management practices. Indeed, the anthropic events on the grasslands, like mowing and/or casual grazing, disturb the natural cycle and the structure of the vegetation. Therefore, it is essential to identify the management practices in each parcel in order to predict their effect on biodiversity and related ecosystem services. In this context, remote sensing appears to be an adapted tool to characterize grasslands at the landscape scale, because of their large spatial coverage and their revisit frequency. However, the reflected signal of the grasslands species mix and diversity is more difficult to interpret compared to mono-specific lands like crops. Furthermore, grasslands are relatively small elements of the landscape (in the order of the hectare), which require high spatial resolution data to be distinguishable. Given their phenological cycle and the punctuality of the anthropic event (i.e., mowing), very dense time series through the vegetation cycle are necessary to identify the management types. Until recently, satellites mission offering high frequency of revisit had low spatial resolution (i.e., MODIS), and high spatial resolution mission did not offer dense time series. Now the sensor capacities have been improved in the field of terrestrial observation, in particular through the ESA Copernicus Programme. Thus, these new missions offer new possibilities for grasslands monitoring. The aim of this study is to identify the management practices of the grasslands through a satellite image time series (SITS). We used a Formosat-2 (8-m spatial resolution) series composed of 17 multispectral images through the year 2013. The study site is located in south-west France, near Toulouse, in a semi-rural area where livestock farming is in decline. The dataset is composed of about 50 parcels with diverse management methods. The management practices were determined from interviewing the farmers. We identified 4 management types during this vegetation cycle: grassland mown once, grassland mown twice, grazing and mixed management (mowing then grazing). We used them as classes for the classification. To remove the noise in the SITS, we smoothed the NDVI of the pixels applying the Whittaker filter. Then we processed a statistical supervised classification using Support Vector Machines. The classification performed well, with an overall accuracy of 71% and a kappa of 0.58. The results were improved by grouping the 4 classes into 2 similar classes: mown grasslands (independent of the number of cuts) and grazed grasslands. The overall accuracy was 82% with a kappa of 0.65. We concluded that it is possible to discriminate the grassland management types with a SITS of 17 images. We suppose that the classification could be improved using a more dense SITS like Sentinel-2 will present. Moreover, we performed the analysis on the NDVI, that is limited to the effect of only 2 bands (Red and NIR). Sentinel-2 will offer more bands, including one in the red-edge that is useful in vegetation characterization. The work will be continued in the aim of determining the mowing and grazing dates. We assume that the future Sentinel-2 time series will allow more precision in the detection of these events and the measurement of their impact on vegetation structure and composition.
Type de document :
Living Planet Symposium, May 2016, Prague, Czech Republic. 2016, 〈http://lps16.esa.int/index.php〉
Liste complète des métadonnées

Littérature citée [3 références]  Voir  Masquer  Télécharger

Contributeur : Stephane Girard <>
Soumis le : vendredi 10 juin 2016 - 09:02:14
Dernière modification le : jeudi 11 janvier 2018 - 06:21:58


  • HAL Id : hal-01326865, version 1



Mailys Lopes, Mathieu Fauvel, Stéphane Girard, David Sheeren, Marc Lang. High Dimensional Kullback-Leibler Divergence for grassland object-oriented classification from high resolution satellite image time series. Living Planet Symposium, May 2016, Prague, Czech Republic. 2016, 〈http://lps16.esa.int/index.php〉. 〈hal-01326865〉



Consultations de la notice


Téléchargements de fichiers