Two Dimensional Histogram based on Relative Entropy Thresholding for Crop Segmentation Using UAV Images

Recently, Unmanned aerial vehicle (UAV) based remote sensing has become a promising way in precision agriculture. Crop or plant segmentation from UAV images plays a vital role in monitoring crop growth. However, the extraction of crops under various illumination conditions is onerous. Numerous methods on segmentation were presented in the literature, out of which threshold-based methods are simple and easy to implement. Previous methods used for crop segmentation utilized complete information of pixels in an image resulting in improper segmentation. The use of local information about pixels can give accurate segmentation. In this work, we constructed a two-dimensional histogram utilizing the gray level of pixels and relative entropy of its neighboring pixels of an contrast enhanced image. The optimal threshold was obtained by minimizing relative entropy criteria. The crops were extracted using logical AND operator on segmented image and a^* channel of CIELAB color space. The proposed method was evaluated on Sorghum and Pearl Millet datasets. The misclassification error, Dice coefficient, Jaccard Index were used to compare the performance of the proposed method, Otsu, and Kapur method. The performance analysis shows that the proposed approach achieved more accurate segmentation than other threshold-based methods.