Design and application of spectrum sensor for weed detection used in winter rape field

Abstract: Due to lack of weed identification and positioning equipment，farmers usually use large area uniform spraying of chemical herbicides, which not only wastes herbicides and labor, but also leads to ecological environment pollution and agricultural product quality problem. At the same time, the weed control accuracy using existing mechanical weed control method is low with a high crop injury rate. Therefore, accurate weed identification is a key issue in target pesticide application and mechanical weed control. There are three kinds of weed identification method: image-based weed identification method, spectrum-based weed identification method, and spectral-image-based weed identification method. At present, spectrum sensor based on spectrum analysis has been most widely accepted in actual weed control due to its advantages of simple system configuration, lossless and high processing speed. Based on the four characteristic wavelengths (595, 710, 755 and 950 nm) selected by the investigation of weeds in the winter rape field, in this paper, we presented our research on weed spectrum sensor. According to the principle of optical system and the actual operation requirements in field, the structure design scheme of the spectrum sensor was proposed, which consisted of five parts, active light source, convex lens, light filter, photocell, and signal conditioning device. As the field measurement results vulnerable to weather conditions, we used LED as an active light source. There were four optical channels (595, 710, 755 and 950 nm) and three active LED light sources in the spectrum sensor. A K9 lenticular lens with diameter of 16 mm and focal length of 16 mm was chosen. A narrow-band interference filter with a center wavelength of 595, 710, 755, 950 nm was applied, whose half-peak bandwidth is 10 nm and aperture is 16 mm. The size of the photocell is 2.65 mm × 2.65 mm. The convex lens, the light filter and the photocell were sequentially arranged in the optical channels in order to detect the spectral information. Signal processing circuits were developed to meet the reliable output signal amplification, filtering and other requirements without distortion. The spectral distance of the spectral sensor is 400-700 mm while the diameter of the field of view is 60-100 mm. After that, the spectral sensor was applied to do calibration test and experimental verification. The calibration test established four mathematical models between the four output results by intelligent spectral sensor and the four measuring results of FieldSpec(r)3 spectrometer. The determination coefficients of each model were 0.799, 0.812, 0.892, and 0.867. The results of the experimental verification showed that most of the relative errors were within 10%, indicating that the designed sensor could separate winter rape from weeds. Therefore, the spectrum sensor could make its contribution to the exploration of weed automatic identification equipment. Experiments on actual weed identification showed that the average recognition rate was 90.7%, which had a good weed recognition effect. The main factors affecting the recognition results were the nature light and the mechanism vibration. How to reduce the influence of interference factors on the recognition precision will be the focus of the next research.