ABSTRACT
Imaging spectrometers enable the calculation of several narrowband vegetation indices and absorption band parameters (e.g., depth, width, area, and asymmetry) associated with vegetation structure, canopy biochemistry and plant physiology. Along with the reflectance and reflectance ratios, these attributes are useful for crop type discrimination using different combinations of metrics and classifiers. This chapter provides an overview of the factors that affect crop type discrimination. By using discriminant analysis and different Hyperion/Earth Observing One (EO-1) images acquired in Brazil, we evaluate the spectral discrimination between flooded rice, coffee, sugarcane, bean, corn, cultivated pasture, and soybean. In addition, we demonstrate the spectral resolution influence on cultivar discrimination from multispectral to hyperspectral remote sensing. Finally, recent advances in feature selection and classification techniques for crop type discrimination are discussed. Overall, the examples presented here show that hyperspectral sensors provide significantly better crop classification results than multispectral sensors. However, the classification performance depends on other factors such as instrumental signal-to-noise ratio (SNR), spectral range of data acquisition, phenological stages of the multiple crops and cultivars, and the adequate use of feature selection and classification techniques. After feature selection to avoid the Hughes effect and reduce data dimensionality, the combination of different metrics increased crop classification accuracy. Depending on the classifier, the use of the whole set of highly correlated hyperspectral metrics may actually produce worse classification results than the use of a selected set of metrics with good discriminatory power. Because the forthcoming spaceborne imaging spectrometers will acquire images with much better SNR and larger swath width than those of Hyperion, the discriminatory power of the metrics and classifiers will be probably much better than that observed in the present study.
