Weed detection with Unmanned Aerial Vehicles in agricultural systems

Organised by Laboratory of Geo-information Science and Remote Sensing

Wed 17 December 2014 09:00 to 09:30

Venue Lumen, gebouwnummer 100
Room 1

By Thomas Koot (The Netherlands)


This study investigated the potential of using Unmanned Aerial Vehicles (UAVs) in an agricultural weed control system. To do so an UAV is deployed to make aerial RGB and hyperspectral images of three agricultural fields planted with sugar beet plants and volunteer potato plants. Those images are geo-referenced and the hyperspectral images go through several spectral and geometrical pre-processing steps after which they are compressed to multispectral images containing five bands: 440 nm – 510 nm (blue), 520 nm – 590 nm (green), 630 nm – 685 nm (red), 690 nm – 730 nm (red-edge), 760 nm – 850 nm (near infra-red).

Both RGB images and multispectral images have been used to train and validate four classification algorithms: one based on an index (greenness and vegetation for respectively RGB images and multispectral images) and three machine learning techniques: Linear discriminant analysis, quadratic discriminant analysis, Artificial Neural Network (ANN).

The results of the training and validation show that the best classification accuracy (99%) is achieved by an ANN when it is validated on the same field it has also been trained on. When the same classification algorithm is then validated for the two other fields, the classification accuracy drops to 71% and 75%. The same pattern is present for the other tested classification algorithms. This pattern shows that the used classification algorithms are condition sensitive and therefore perform much better on fields they have been trained on than on other fields with the same plants but recorded under different conditions.

Using RGB images as input outperformed the classification where multispectral images were used as input for all tested classification algorithms. However some concerns have been raised on the band choices for the creation of the multispectral images and the use of those bands. Firstly the bandwidths and ranges could have been chosen based on an optimal signal-to-noise ratio and secondly the red-edge band could have been used as a specific wavelength where the red-edge occurs rather than the average reflection value in the red-edge range.

Some propositions are made for a weed control system in which an UAV assists an weed detecting and removing Unmanned Ground Vehicle (UGV) in two different ways: 1. triggering the UGV to go into the field; 2. provide the UGV with an optimum path through the field.

Keywords: weed detection; UAV; crop classification; machine learning; automated weed control