Agritrop
Accueil

Full-polarization radar remote sensing and data mining for tropical crops mapping: a successful SVM-based classification model

Denize Julien, Todoroff Pierre, Le Mezo Lionel, Corgne Samuel. 2015. Full-polarization radar remote sensing and data mining for tropical crops mapping: a successful SVM-based classification model. . San Francisco : AGU, Résumé, 1 p. AGU Fall Meeting, San Francisco, États-Unis, 14 Décembre 2015/18 Décembre 2015.

Communication par affiche
[img]
Prévisualisation
Version publiée - Anglais
Utilisation soumise à autorisation de l'auteur ou du Cirad.
ID591291.pdf

Télécharger (109kB) | Prévisualisation
[img] Version publiée - Anglais
Accès réservé aux agents Cirad
Utilisation soumise à autorisation de l'auteur ou du Cirad.
Poster AGU DENIZE_Julien_B43A-0517.pdf

Télécharger (6MB) | Demander une copie

Matériel d'accompagnement : 1 poster

Résumé : In Reunion, a tropical island of 2,512 km², 700 km east of Madagascar in the Indian Ocean, constrained by a rugged relief, agricultural sectors are competing in highly fragmented agricultural land constituted by heterogeneous farming systems from corporate to small-scale farming. Policymakers, planners and institutions are in dire need of reliable and updated land use references. Actually conventional land use mapping methods are inefficient under the tropic with frequent cloud cover and loosely synchronous vegetative cycles of the crops due to a constant temperature. This study aims to provide an appropriate method for the identification and mapping of tropical crops by remote sensing. For this purpose, we assess the potential of polarimetric SAR imagery associated with associated with machine learning algorithms. The method has been developed and tested on a study area of 25*25 km thanks to 6 RADARSAT-2 images in 2014 in full-polarization. A set of radar indicators (backscatter coefficient, bands ratios, indices, polarimetric decompositions (Freeman-Durden, Van zyl, Yamaguchi, Cloude and Pottier, Krogager), texture, etc.) was calculated from the coherency matrix. A random forest procedure allowed the selection of the most important variables on each images to reduce the dimension of the dataset and the processing time. Support Vector Machines (SVM), allowed the classification of these indicators based on a learning database created from field observations in 2013. The method shows an overall accuracy of 88% with a Kappa index of 0.82 for the identification of four major crops.

Mots-clés libres : Radar remote sensing, Crop mapping, Polarimetry, Classification, SVM

Auteurs et affiliations

  • Denize Julien, Université de Rennes 2 (FRA)
  • Todoroff Pierre, CIRAD-PERSYST-UPR AIDA (REU)
  • Le Mezo Lionel, CIRAD-PERSYST-UPR AIDA (REU)
  • Corgne Samuel, Université de Rennes 2 (FRA)

Source : Cirad-Agritrop (https://agritrop.cirad.fr/591291/)

Voir la notice (accès réservé à la Dist) Voir la notice (accès réservé à la Dist)

[ Page générée et mise en cache le 2020-06-17 ]