Crop/Weed Discrimination Using a Field Imaging Spectrometer System | JoVE Visualize

Area of Science:

Agricultural Science
Remote Sensing
Spectroscopy

Background:

Sensors are crucial for smart agriculture, with spectroscopy showing promise for weed detection.
Existing research often lacks field applicability, focusing on controlled environments.
Hyperspectral imaging generates large datasets requiring efficient dimensionality reduction and feature extraction.

Purpose of the Study:

To design and utilize a Field Imaging Spectrometer System (FISS) for discriminating crops and weeds in real-world field conditions.
To evaluate dimensionality reduction techniques, specifically wavelet transform, for hyperspectral data processing in weed detection.
To compare the performance of wavelet coefficients versus raw spectral bands as classification features and assess different discrimination methods.

Main Methods:

A Field Imaging Spectrometer System (FISS) operating from 380-870 nm with 344 bands was employed.
Wavelet transform was used for dimensionality reduction, with extracted coefficients serving as classification features.
Feature selection utilized Wilks' statistic-based stepwise selection, and classification was performed using Fisher's linear discriminant analysis (LDA) and Support Vector Machine (SVM).

Main Results:

High classification accuracy (>85%) was achieved using a limited number of spectral bands (8), improving to >90% with 15 bands.
Red edge spectral bands demonstrated significant discriminant capability for crop-weed differentiation.
Wavelet coefficients generally outperformed raw spectral bands as features, especially with fewer variables, while SVM showed superior performance for nonlinear classification.

Conclusions:

Effective multiclass discrimination of crops and weeds is feasible in field environments using hyperspectral imaging and advanced data processing.
Wavelet transform offers an efficient method for dimensionality reduction in hyperspectral data for weed detection applications.
Support Vector Machine (SVM) is a robust classification method for complex spectral data, outperforming LDA in this study.