Selective application of herbicide to weeds at an earlier stage in crop growth is an important aspect of site-specific
management of field crops. For approaches more adaptive in developing the on-line weed detecting application, more
researchers involves in studies on image processing techniques for intensive computation and feature extraction tasks to
identify the weeds from the other crops and soil background. This paper investigated the potentiality of applying the
digital images acquired by the MegaPlus<sup>TM</sup> MS3100 3-CCD camera to segment the background soil from the plants in
question and further recognize weeds from the crops using the Matlab script language. The image of the near-infrared
waveband (center 800 nm; width 65 nm) was selected principally for segmenting soil and identifying the cottons from
the thistles was achieved based on their respective relative area (pixel amount) in the whole image. The results show
adequate recognition that the pixel proportion of soil, cotton leaves and thistle leaves were 78.24%(-0.20% deviation),
16.66% (+ 2.71% SD) and 4.68% (-4.19% SD). However, problems still exists by separating and allocating single
plants for their clustering in the images. The information in the images acquired via the other two channels, i.e., the
green and the red bands, need to be extracted to help the crop/weed discrimination. More optical specimens should be
acquired for calibration and validation to establish the weed-detection model that could be effectively applied in fields.
The traditional uniform herbicide application often results in an over chemical residues on soil, crop plants and
agriculture produce, which have imperiled the environment and food security. Near-infrared reflectance spectroscopy
(NIRS) offers a promising means for weed detection and site-specific herbicide application. In laboratory, a total of 90
samples (30 for each species) of the detached leaves of two weeds, i.e., threeseeded mercury (<i>Acalypha australis </i>L.) and
fourleafed duckweed (<i>Marsilea quadrfolia </i>L.), and one crop soybean (<i>Glycine max</i>) was investigated for NIRS on 325-
1075 nm using a field spectroradiometer. 20 absorbance samples of each species after pretreatment were exported and
the lacked Y variables were assigned independent values for partial least squares (PLS) analysis. During the combined
principle component analysis (PCA) on 400-1000 nm, the PC1 and PC2 could together explain over 91% of the total
variance and detect the three plant species with 98.3% accuracy. The full-cross validation results of PLS, i.e., standard
error of prediction (SEP) 0.247, correlation coefficient (<i>r</i>) 0.954 and root mean square error of prediction (RMSEP)
0.245, indicated an optimum model for weed identification. By predicting the remaining 10 samples of each species in
the PLS model, the results with deviation presented a 100% crop/weed detection rate. Thus, it could be concluded that
PLS was an available alternative of for qualitative weed discrimination on NTRS.
This research aimed to identify weeds from crops in early stage in the field operation by using image-processing
technology. As 3CCD images offer greater binary value difference between weed and crop section than ordinary digital
images taken by common cameras. It has 3 channels (green, red, ifred) which takes a snap-photo of the same area, and
the three images can be composed into one image, which facilitates the segmentation of different areas. By the
application of image-processing toolkit on MATLAB, the different areas in the image can be segmented clearly. As edge
detection technique is the first and very important step in image processing, The different result of different processing
method was compared. Especially, by using the wavelet packet transform toolkit on MATLAB, An image was
preprocessed and then the edge was extracted, and getting more clearly cut image of edge. The segmentation methods
include operations as erosion, dilation and other algorithms to preprocess the images. It is of great importance to
segment different areas in digital images in field real time, so as to be applied in precision farming, to saving energy and
herbicide and many other materials. At present time Large scale software as MATLAB on PC was used, but the
computation can be reduced and integrated into a small embed system, which means that the application of this
technique in agricultural engineering is feasible and of great economical value.
This research aimed to identify weeds from crops in early stage in the field by using image-processing technology. As 3CCD images offer greater binary value difference between weed and crop section than ordinary digital images taken by common cameras. It has 3 channels (green, red, ir red), which takes a snap-photo of the same area, and the three images can be composed into one image, which facilitates the segmentation of different areas. In this research, MS3100 3CCD camera is used to get images of 6 kinds of weeds and crops. Part of these images contained more than 2 kinds of plants. The leaves' shapes, sizes and colors may be very similar or differs from each other greatly. Some are sword-shaped and some (are) round. Some are large as palm and some small as peanut. Some are little brown while other is blue or green. Different combinations are taken into consideration. By the application of image-processing toolkit in MATLAB, the different areas in the image can be segmented clearly. The texture of the images was also analyzed. The processing methods include operations, such as edge detection, erosion, dilation and other algorithms to process the edge vectors and textures. It is of great importance to segment, in real time, the different areas in digital images in field. When the technique is applied in precision farming, many energies and herbicides and many other materials can be saved. At present time large scale softwares as MATLAB on PC are also used, but the computation can be reduced and integrated into a small embedded system. The research results have shown that the application of this technique in agricultural engineering is feasible and of great economical value.
It is difficult to measure the leaf area of plants by conventional methods as a result of the irregular leaf shapes. In this research, a new method was presented to measure leaf area using image processing techniques. The leaves images were acquired by MS3100 3CCD camera, and each image was composed of three channel data (green, red, near-infrared). The image data were transferred to a host computer and were stored as files in TIFF or JPEG format. Some image processing methods were applied to calibrate the leaf image, detect the margin of the leaf, and calculate the area by counting the pixels in the leaf. From the experimental results, it shows that the image method has good measurement accuracy; the relative error is less than 0.5 percent; and image processing is a rapid and non-destructive tool to measure the leaf area of plants.