Enhanced Tree-Based Multiple Classifier Systems for Cropland Classification from Optical and Full-Polarimetric Radar Images

Optical and polarimetric synthetic aperture radar (PolSAR) earth observations offer valuable sources of information for agricultural applications and crop mapping. Various spectral features, vegetation indices and textural indicators can be extracted from optical data. These features contain information about the reflectance and spatial arrangement of crop types. By contrast, PolSAR data provide quad-polarization backscattering observations and target decompositions, which give information about the structural properties and scattering mechanisms of different crop types. Combining these two sources of information can present a complementary data set with a significant number of spectral, textural, and polarimetric features for crop mapping and monitoring. Moreover, a temporal combination of both observations may lead to obtaining more reliable results compared to the use of single-time observations. However, there are several challenges in cropland classification using this large amount of information. The first challenge is the possibility correlation among some optical features or radar features which leads to redundant features. Moreover, some optical or radar features may have a low relevancy with some or all crop types. These two challenges cause to increase complexity and computational load of classification. In addition, when the ratio of number of samples to the number of features is very low, the curse of dimensionality may be occur. Another challenge of classification is the imbalanced distribution among various crop types, the so called imbalanced data. Various classifier have been presented for cropland classification from optical and radar data. Among these classifiers, the multiple classifier systems (MCS) especially the random forest (RF). The main aim of this paper is an alternative to RF which is able to solve these two challenges, the curse of dimensionality and the imbalanced data, simultaneously. The proposed MCSs have other modifications in feature selection and fusion steps of RF. These two methods called as balanced filter forest (BFF) and cost-sensitive filter forest (CFF).
The study area of this paper was the southwest district of Winnipeg, Manitoba, Canada, which is covered by various annual crops. The data used in this paper were bi-temporal optical and radar images acquired by RapidEye satellites and the UAVSAR system. RapidEye is a spaceborne satellite, which has five spectral channels: blue (B), green (G), red (R), NIR and RE. In this paper, two optical images were collected on 5 and 14 July 2012. Both these images were orthorectified on the local North American 1983 datum (NAD-83) with a spatial resolution of about 5 m. The UAVSAR sensor is an airborne SAR sensor, which operates in the L-band frequency in full polarization mode (i.e., HH, HV, VH and VV). The radar images used in this paper were simultaneously acquired with the optical images. They were orthorectified on the World Geodetic System 1984 datum (WGS-84) with an SRTM3 digital elevation model. They were also multilooked by 2 pixels in azimuth and 3 pixels in range directions. Moreover, the de-speckling process, using a 5 × 5 boxcar filter, was applied to the data in order to alleviate the speckle effect. The spatial resolution of these images was then approximately 15 m.
The results indicated that the proposed methods could increase the overall accuracy to 10% and the speediness to 6 times more than the classical RF method.