[Identification methods of crop and weeds based on Vis/NIR spectroscopy and RBF-NN model].

The automated recognition of crop and weed by using Vis/NIR spectral in field is one of hottest research branches of agriculture engineering. If the recognition is efficient and effective, then the variate operations of herbicide or fertilizer spraying in field could be realized. Many researches have pointed out that the reflectance rate of green plant leaves could be used to identify the varieties. As the colors and surface textures of crop and weed were change in different living phases, these changes may exert great influence on the reflectance spectral of plant leaves. Vis/NIR spectra of three weeds and one crop in two different terms were recorded by spectral meter ASD FieldSpec Pro FR. Its wave band is from 325 to 1 075 nm. The scan time was 270 ms. The scanning times of per sample was set to 30 times. Firstly, 23 days after the planting of soybean, some soybean leaves and weeds leave were picked from the field, and brought to lab to record spectral. The lighting condition was controlled by an artificial halogen bulb. Secondly, on the 45th day, the same experiment was done. The three weeds were goose grass, alligator alternanthera and emarginate amaranth. The crop was soybean seedling. Totally 378 samples were drawn for two terms. As one original reflectance spectrum contains 651 float numbers, the total data volume was huge. Using wavelet transform to compress data volume and extract characteristic spectral data was tried. The result was 114 float numbers per sample. Among them, 250 samples from two terms were used as input data to build artificial neural network model, and those left were used to check the validation. Radial basis function neural network model is widely used in pattern recognition problems. It is a nonlinear and self adaptive parallel. By assigning a 1 by 4 vector to each spectral samples, the source data could be used to build an RBF-NN model. All the samples were assigned these standard output data. Then, the left 128 samples were used to check the performance of the model. The result is that only 3 samples from the second term of goose grass were wrongly classified as alligator alternanthera, which showed that RBF neural network have strong ability to differentiate spectra of species of plant, and that there was no massive difference of NIR spectra of one plant in different life periods. This demonstrated that the NIR spectra could be used to identify crop from weed with no need to care about the living stages of these plants.