GEOSCAN, résultats de la recherche


TitreAn improved scheme for rice phenology estimation based on time-series multispectral HJ-1A/B and polarimetric RADARSAT-2 data
AuteurYang, Z; Shao, Y; Li, K; Liu, Q; Liu, L; Brisco, B
SourceRemote Sensing of Environment vol. 195, 2017 p. 184-201,
Séries alt.Ressources naturelles Canada, Contribution externe 20181304
ÉditeurElsevier BV
Documentpublication en série
Mediapapier; en ligne; numérique
Sujetstélédétection; géophysique
ProgrammeGéosciences de changements climatiques
Diffusé2017 04 25
Résumé(disponible en anglais seulement)
Rice phenology information is critical for farm management and productivity evaluation. Synthetic aperture radar (SAR) and optical remote sensing data are very useful for monitoring rice growth. Therefore, this paper focuses on building a robust, accurate and operational scheme for phenology estimation based on SAR and multispectral remote sensing data. The proposed scheme has three main improvements. First, two main types of rice field - transplanted indica rice field (TRF) and direct-sown japonica rice field (DRF) - were considered individually. Second, 149 SAR signatures and optical Vegetation Indices (VIs) were extracted from time-series of optical and full/compact-pol SAR data and a novel feature selection method based on Monte Carlo experiments and the correlation limitation (MCCL) was proposed. Third, six different phenological labeling cases were considered, with the multiclass relevant vector machine (mRVM) being used to identify different stages. Based on the improvements above, we generated the optimal feature matrix based on the MCCL; this matrix consisted of the optimal feature subset (OFS) used for the identification of each stage. The overall phenological estimation accuracy (OPEA) based on the optical VIs and SAR signatures together (86.59%), was shown to be better than that based on either the optical VIs or the SAR signatures only. In addition, it was significant to consider the differences between the DRF and TRF for rice phenology estimation. The OPEA was higher than 85% when the DRF and TRF were considered separately, much higher than the accuracy obtained (68.89%) when the differences between the DRF and TRF were ignored. Lastly, the availability of the proposed feature selection method MCCL was discussed. It was found that the best OPEA was generated using the OFS selected by the proposed MCCL when the dimensions of the OFS were kept small.