By Vânia L.C.A. Souza:
Summary
Mapping accurately agricultural fields is one of the main challenges for monitoring areas using remote sensing. Crops change during the growing season and it is often desirable to use images acquired at several dates for plant identification. However, there is a high cost for these image acquisitions. Airborne or satellite campaigns usually cover specific regions for specific dates then offering limited data. In the past, studies on crop classification were mostly performed using satellite multispectral sensors. A multispectral sensor provides data over a large area gathered quickly and economically from satellite platforms but with small spectral resolution. Advances in sensor technologies created hyperspectral sensors to overcome this spectral limitation of multispectral sensors. A hyperspectral sensor collects spectral data in several hundreds of bands in a simple acquisition time offering opportunity to discriminate more precisely different plants. Thus, there is a chance that hyperspectral data produce better crop classification especially if it is combined with multiview angle information. The purpose of this study was to evaluate the possible improvement of the accuracy of crop classification by using hyperspectral and directional remote sensing data. It investigated effects of number of bands, bandwidth and bidirectional information on classification. Four AHS-160 hyperspectral images were used in this research. They covered areas of barren land, barley, beet, grass, horticulture, maize, onion, potato and wheat located in the region of the Gelderse Poort in the Netherlands. Classification accuracy was investigated using Maximum Likelihood with ground truth data. Analyses were divided in two parts: In the first part the effect of bandwidth and number of bands was investigated by comparing classification results of a hyperspectral image and a simulated multispectral image. Results indicated that the size of bandwidth did not affect classification. An image with a narrow band and an image with a broad band had not a statistically different classification. The effects of the number of bands were analyzed comparing the classification of an image with 63 bands with an image with 7 bands. The image with 63 bands had better user and producer accuracy of crops than the image with 7 bands. The second part studied effects of multi view angle or bidirectional reflectance on classification. A common area in various images was considered in this part of study. Classification results using one image in one flight direction was compared with an overlap image that combined three flights directions. A principal component analysis was applied to reduce dimensionality of 189 bands of the overlap image. Seven principal components of the overlap image had better classification results compared with single images. In conclusion, Bidirectional reflectance increased classification accuracy of the majority of crops in hyperspectral images. The bidirectional reflectance influenced positively classification results of crops specially located in the principal plane.
Key words: hyperspectral; bidirectional reflectance; viewing geometry; narrow bands; broad band; number of bands; crop classification; principal components analyses.