Towards the characterization of crop and weeds at leaf scale: A large comparison of shape, spatial and textural features

Abstract

In the context of crop and weeds discrimination, different methods are used to detect and classify plants from an acquisition system. Various estimators and descriptors are commonly used to characterize plants within an image. However, the available studies are based on disparate criteria, plants, and acquisition materials which does not allow an accurate estimation of the potential of criteria combinations applied to a new study. Thus, the objective of this study is to: (1) experimentally evaluate the discrimination potential of each criterion at the leaf scale, using images taken in field condition; (2) optimize the parameters of these criteria; and (3) determine the best combination of criteria to use.A literature review is conducted to determine the set of criteria that could be used. A set of 3545 criteria is studied with an algorithm defined to select the best subsets of features (evaluated on a ground truth dataset). Finally, a classification of the vegetation cover is proposed, using the best performing subset. Results show the importance of selecting a smaller set of properties (at most 20 features among the 3545 available) and associating different feature types (for instance spatial with textural and morphological features).

Keywords

• Leaf classification
• Crop and weed discrimination
• Remote sensing
• Multispectral