Identification method of multi-feature weed based on multi-spectral images and data mining

Abstract: Field weed detection is one of the key problems in realizing the variable precision applying pesticide to take place of the herbicide. Image-based weed classification and spectral information of plants are useful to detect weeds in real-time using multi-spectral features. Aimed to meet the identification accuracy requirements of variable spraying on weed, a new method using decision tree algorithm-C4.5 of data mining was developed to discriminate or classify crop and weeds by the multi-spectral images. The multi-spectral images of weeds and maize were captured by MS4100 Duncan Camera in the test field of Northwest Agriculture and Forestry University on May, 2012, and transformed from CIR color space to Lab systems, which can distinguish different quantized color and measure the Euclidean distance of different colors. Then vegetation was segmented from soil using K-means clustering algorithm. Mathematical morphology was used to fill small holes among the extracted vegetation leaves, and connect the uncompleted contour line of the discontinuous edges which may be caused by noise, occlusion and other factors. Contour tracing was used to get the contours of leaves. After these image processing, shape features, texture features and fractal dimensions of the vegetation were extracted. A random sample of 120 images from all 240 images were involved in this study as the training samples, 20 images from 40 images were used as the test samples. The results of statistic analysis showed that multi-feature combining with shape feature, texture feature and fractal dimension together achieved the highest recognition rate of 96.3%, compared to the single feature recognition rate of 75.0%. To validate the feasibility of this study, C4.5 algorithm was compared with BP (error back propagation) algorithm and SVM (support vector machine) algorithm in recognizing multi-feature. The experimental results showed the average recognition rates were 92.5% and 95.0% for BP and SVM algorithms, respectively. The results showed that the average recognition rate of C4.5 algorithm was higher than that of the other two algorithms and it was an effective and feasible method to rapidly identify the weeds. The results provide a technical basis for accurate spraying on corn seedling. Further studies could be conducted in weed recognition, such as testing robust algorithms in the real complex environment (e.g. the uneven illumination, random distribution of the vegetation growing position), and discussing decision level fusion of feature data to further reduce the dimensions of data.