Crop Classification Based on Feature Band Set Construction and Object-Oriented Approach Using Hyperspectral Images

Remote sensing plays a significant role for crop classification. Accurate crop classification is a common requirement to precision agriculture, including crop area estimation, crop yield estimation, precision crop management, etc. This paper developed a new crop classification method involving the construction and optimization of the vegetation feature band set (FBS) and combination of FBS and object-oriented classification (OOC) approach. In addition to the spectral and textural features of the original image, 20 spectral indices sensitive to the vegetation's biological parameters are added to the FBS to distinguish specific vegetation. A spectral dimension optimization algorithm of FBS based on class-pair separability (CPS) is also proposed to improve the separability between class pairs while reducing data redundancy. OOC approach is conducted on the optimized FBS based on CPS to reduce the salt-and-pepper noise. The proposed classification method was validated by two airborne hyperspectral images. The first image acquired in an agricultural area of Japan includes seven crop types, and the second image acquired in a rice breeding area consists of six varieties of rice. For the first image, the proposed method distinguished different vegetation with an overall accuracy of 97.84% and kappa coefficient of 0.96. For the second image, the proposed method distinguished the rice varieties accurately, achieving the highest overall accuracy (98.65%) and kappa coefficient (0.98). Results demonstrate that the proposed method can significantly improve crop classification accuracy and reduce edge effects, and that textural features combined with spectral indices sensitive to the chlorophyll, carotenoid, and Anthocyanin indicators contribute significantly to crop classification. Therefore, it is an effective approach for classifying crop species, monitoring invasive species, as well as precision agriculture related applications.