Automated detection and control of volunteer potato plants

High amounts of manual labor are needed to control volunteer potato plants in arable fields. Due to the high costs, this leads to incomplete control of these weed plants, and they spread diseases like Phytophthora infestans to other fields. This results in higher environmental loads by curative spraying of crop protection chemicals, which is in contradiction to the required decreased use of crop protection chemicals to save the environment. Therefore, the main objective of this thesis was “to develop a system for automated detection and control of volunteer potato plants”. A systematic design approach was used to define a program of requirements and to identify and order possible solutions to accomplish the detection and control. The main requirements were a travel speed of up to 2 m s-1, resolution of control at least 10×10 mm, work under variable natural light conditions, control of volunteer plants > 95%, and undesired control of sugar beet plants < 5%. The design strategy resulted in color and near-infrared machine vision as detection method and a micro-sprayer for application of glyphosate as a result. Furthermore, issues were identified that required further investigation to successfully come to a proof of principle machine. The research was then focused on:
- Detection of volunteer potato plants,
- Control of volunteer potato plants,
- Real-time implementation of integrated detection and control on a proof of principle machine.

For the purpose of detection of volunteer potato plants, the narrow band spectral reflectance properties of volunteer potato plants and sugar beet plants were analyzed. Narrow band spectral measurements were done in 2006 and 2007 on two different fields. This resulted in 15 datasets on clay and sand soil. Discriminating wavebands were selected and classified with neural networks and statistical discriminant analysis. A neural network with two hidden neurons performed best for classification. Two sensors were used covering the range from 450 to 900 nm and from 900 to 1650 nm. Both visible and near infra-red wavebands were responsible for discrimination. From the analysis 450, 765, and 855 nm from sensor 1 and 900, 1440, and 1530 nm from sensor 2 were identified as important discriminative wavebands. However, the discriminative wavelengths depended on field and crop status and could not be generalized. Ten wavebands that were optimally adapted to the datasets gave 99% true negative classification of volunteer potato plants. On the other hand, a fixed set of three wavebands that was not adapted to the individual datasets gave 80% true negative classification of volunteer potato plants. This indicates that adaptive feature sets are required for classification.

The development of the machine vision detection system started with measurements in 2005. Color based detection showed that the difference in classification results was larger between fields than the difference between a static neural network and static Bayesian classification. Then, machine vision measurements in 2006 with a color camera under changing and constant natural light conditions showed that crop and weed properties change within a field. An adaptive instead of static classification increased classification accuracy from 34.9% to 67.7% under changing light conditions. Under constant natural light conditions, the classification accuracy increased from 84.6% to 89.8%. So, adaptive classifiers are required and were implemented in the further research as these gave significantly higher classification results. As a next step, besides a color camera also a near-infrared camera was used for imaging within the proof of principle machine, as this gave a better feature set for classification. Additionally, the field of view of the cameras was shielded and artificial light was used to maintain constant light conditions. For the real-time implementation, an unsupervised adaptive Bayesian classifier was used. The crop row position and crop row width were determined and a Kalman filter improved tracking of the rows, to adapt to the varying properties of the crop in the field. Data from between the crop rows was trained as the volunteer potato class and data from within the crop row was trained as the sugar beet class. This resulted in good quality training data for the Bayes classifier. The system was unsupervised, as it learned and trained itself based on row recognition. The features that were used for training and classification were: blue, hue, saturation, excessive green, red minus blue, near-infrared and near-infrared difference vegetation index (NDVI). These feature values within the training data were continuously locally adapted, in two first-in-first-out buffers both with an area of 500 cm2 for sugar beet and volunteer potato plants. Measurements were done on seven days in 2007 and 2008. The results showed a trade-off between the percentage of correct classified volunteer potato plants and the percentage of misclassification of sugar beet plants. In one of the fields 96.6% volunteer potato classification and 8.0% sugar beet misclassification was achieved.

Connected to the detection system was a micro-sprayer that applied glyphosate in gel to the volunteer potato plants. Spraying gel through a micro-sprayer was innovative. This proved to work in the application of glyphosate on plants. As knowledge of the dose response of glyphosate on potato was outdated and could not be used for plant specific application, a dose-response study was done with flat fan nozzles on 120 potato plants to determine the efficacy of glyphosate. The effect parameters tuber weight and photosynthesis activity were analyzed with log-logistic nonlinear regression methods. This resulted in an amount of 843 μg a.e. per plant for reduction of tuber weight and photosynthesis with 90%. This amount was applied on plants with a height of 6.1±1.39 cm and an area of 53.3±19.6 cm2. As glyphosate was to be applied with a micro-sprayer, the dose-response study was extended to 500 greenhouse grown potato plants. Five application methods were used: 1) flat fan water application, 2) flat fan gel application, 3) micro-sprayer low density distribution, 4) micro-sprayer medium density distribution, and 5) micro-sprayer high density distribution. As effect parameters again tuber weight, photosynthesis activity, and in addition shoot dry weight were used. They were analyzed with ANOVAs and box-plots. The micro-sprayer dense distribution with 3022 droplets m-2 and 3.3 mg per droplet had the best efficacy. The micro-sprayer controlled the volunteer potato plants with less glyphosate compared to flat fan nozzles. Furthermore, it had a centimeter precision resolution and low risks of unwanted crop damage.

With real-time hardware, machine vision detection and micro-sprayer were integrated to a proof of principle machine. A travel speed of 0.8 m s-1 was reached with the proof of principle machine and it had an approximated capacity of 2.5 hrs ha-1. This was the maximum that could be realized as the micro-sprayer valve actuation frequency was maximally 80 Hz. The image processing time for one image of 0.2 m length was 195 ms. At this travel speed automated feedback systems on the operation of the system are required to support and replace human surveillance. Therefore, the Fréchet distance measure between multivariate distributions was introduced as quality indicator of classification performance. The Fréchet distance measure was significantly smaller when the classification performance was low, as identified on ground truth determined classification results afterwards. This proves that the performance could be predicted with a distance measure between multivariate distributions. In case of poor predicted classification performance, the application of glyphosate with the micro-sprayer can be halted to prevent unwanted crop damage and economic losses. The accuracy of application was ±1.4 cm in longitudinal direction and ±0.75 cm in transversal direction. During a field trial, up to 84% of the volunteer plants were controlled with 1.4% unwanted controlled sugar beet plants.

To sum up, within this research a proof of principle machine for automated detection and control of volunteer potato plants in sugar beet fields has successfully been developed. The system performed closely to the requirements that were set in the start-up of the project. The percentage of 95% controlled volunteer potato plants can be reached. On the other hand, the travel speed still has to be increased from 0.8 m s-1 to 2.0 m s-1. The system is an example of new technology that can be developed for practical applications to reduce the amount of required labor and to reduce the crop protection inputs for weed control in arable farming.