Search

Rumex obtusifolius is a common grassland weed that is hard to control in a non-chemical way. The objective of our research was to automate the detection of R. obtusifolius as a step towards fully automated mechanical control of the weed. We have developed a vision-based system that uses textural analysis to detect R. obtusifolius against a grass background. Image sections are classified as grass or weed using 2-D Fourier analysis. We conducted two experiments. In the first (laboratory) experiment, we collected 28 images containing R. obtusifolius and 28 images containing only grass. Between 23 and 25 of 28 images were correctly classified (82–89%) as showing R. obtusifolius; all grass images were correctly classified as such. In the second (field) experiment, a self-propelled platform was used to obtain five sequences of images of R. obtusifolius plants. We used the parameters that gave the best classification results in the first experiment. We found, after changing one of the algorithm’s parameters in response to prevailing light conditions, that we were able to detect R. obtusifolius in each image of each sequence. The algorithm scans a ground area of 1.5 m2 in 30 ms. We conclude that the algorithm developed is sufficiently fast and robust to eventually serve as a basis for a practical robot to detect and control R. obtusifolius in grassland. [ABSTRACT FROM AUTHOR]

For crop management, information on the actual status of the crop is important for taking decisions on nitrogen supply, water supply or harvesting. One would also like to take into account the local spatial variation of the crop. Remote sensing has proved to be a useful technique for estimating and mapping the spatial variation of various biophysical variables. Calibration of the image data is crucial in the performance and applicability of this technique. The aim of this paper is to show the possibility to calibrate remotely sensed imagery using fast and non-destructive close-range (below 1.3 m height) sensing instruments, thus providing a means for the assessment of plant characteristics over large areas at low costs. This concept was tested on a homogeneously managed grassland field, subdivided into 20 plots of 15×3 m, at the end of July 2004. Reflected radiation was recorded with an active close-range sensing device, consisting of a visible light and near-infrared (NIR) imaging spectrograph, and a 3CCD camera, equipped with special band filters (central wavelengths are at 600, 710 and 800 nm). An airborne campaign with a four-band UltraCam digital CCD camera was used for extrapolation to larger scales. Plots were harvested, and fresh and dry biomass and leaf nitrogen content were determined. Partial least squares (PLS) models combining spectral and spatial information from the close-sensing device yielded acceptable results in predicting grassland yields and nitrogen content. Subsequently, these predictions were used to calibrate a model with the image data of the remote sensing device. These were then compared, using leave-one-out cross-validation, with the measured field variables, and the model proved to have an acceptable predictive power. [ABSTRACT FROM AUTHOR]