GEOSCAN Search Results: Fastlink


TitleRandom forest classification of crops using quad-polarization synthetic aperture radar (SAR)
AuthorHong, G; Wang, S; Li, J
SourceRemote Sensing of Environment .
Alt SeriesEarth Sciences Sector, Contribution Series 20140303
ProgramAquifer Assessment & support to mapping, Groundwater Geoscience
AbstractPolarimetric SAR is widely used in crop type mapping since it is weather-independent, complementary to optical sensors, and sensitive to crop canopy structure and density. Non-parametric classifiers are usually used in polarimetric SAR classification since they are not limited to a specific statistical distribution as conventional statistical-based classifiers. Random Forest (RF), a non-parametric classifier, has been increasingly used in remote sensing classification since it is an ensemble of tree classifiers using bootstrap aggregated sampling to build individual decision tree and its performance is superior to many other tree-based classifiers. This study proposed RF for crop type classification with a coherency matrix of quad-polarization SAR as the input. The performance of RF was assessed by comparing with two other advanced non-parametric classifiers: support vector machine (SVM) and neural network (NN). Two other inputs based on polarimetric decompositions, Cloude and Pottier and Freeman¿Durden decompositions, were also used to examine the effects of inputs on different classifiers. The results indicated that RF achieved similar accuracy to SVM and both were higher than NN. RF has the advantages of being easier and faster to operate as it only needs two input parameters and does not require parameter tuning. Comparisons of different inputs showed that the coherence matrix outperformed the other two polarimetric decomposition inputs for all classifiers.
Summary(Plain Language Summary, not published)
Mapping vegetation cover is an imperative step to map the surface water cycle and better understand the groundwater recharge. This paper proposed the Random Forest (RF) for mapping crop types using Quad-Polarization Synthetic Aperture Radar (SAR) data from the Radarsat-2 satellite. The performance of RF was assessed by comparing with two other advanced classifiers: support vector machine (SVM) and neural network (NN). The results indicated that RF outperforms other approaches and has the advantages of being easier and faster to operate, and needs less parameter inputs and no parameter tuning. Tests of different inputs for the above classifiers indicated that the coherence matrix has the best performance for all classifiers. The results will improve the accuracy for crop mapping, provide better quality data for water cycle mapping and water resources assessment using the ESS EALCO modelling system, and ensure that satellite data products are improved and used in ESS programs to address the NRCan mandates.