Crop/Weed Discrimination Using a Field Imaging Spectrometer System

Abstract

Nowadays, sensors begin to play an essential role in smart-agriculture practices. Spectroscopy and the ground-based sensors have inspired widespread interest in the field of weed detection. Most studies focused on detection under ideal conditions, such as indoor or under artificial lighting, and more studies in the actual field environment are needed to test the applicability of this sensor technology. Meanwhile, hyperspectral image data collected by imaging spectrometer often has hundreds of channels and, thus, are large in size and highly redundant in information. Therefore, a key element in this application is to perform dimensionality reduction and feature extraction. However, the processing of highly dimensional spectral imaging data has not been given due attention in recent studies. In this study, a field imaging spectrometer system (FISS; 380–870 nm and 344 bands) was designed and used to discriminate carrot and three weed species (purslane, humifuse, and goosegrass) in the crop field. Dimensionality reduction was performed on the spectral data based on wavelet transform; the wavelet coefficients were extracted and used as the classification features in the weed detection model, and the results were compared with those obtained by using spectral bands as the classification feature. The classification features were selected using Wilks’ statistic-based stepwise selection, and the results of Fisher linear discriminant analysis (LDA) and the highly dimensional data processing-oriented support vector machine (SVM) were compared. The results indicated that multiclass discrimination among weeds or between crops and weeds can be achieved using a limited number of spectral bands (8 bands) with an overall classification accuracy of greater than 85%. When the number of spectral bands increased to 15, the classification accuracy was improved to greater than 90%; further increasing the number of bands did not significantly improve the accuracy. Bands in the red edge region of plant spectra had strong discriminant capability. In terms of classification features, wavelet coefficients outperformed raw spectral bands when there were a limited number of variables. However, the difference between the two was minimal when the number of variables increased to a certain level. Among different discrimination methods, SVM, which is capable of nonlinear classification, performed better.