Extraction of cotton seedling growth information using UAV visible light remote sensing images

Abstract: The rapid and accurate seedling situation acquisition is an important prerequisite for farmland fine management, and also the basis for promoting the development of precision agriculture. It was found that UAV remote sensing images combined with machine vision technology had obvious advantages in crop detection in the field. However, current research mainly focused on crops such as corn, wheat, and rape, and only realized the extraction of emergence rate or coverage. In fact, there were few reports on the research of cotton overall seedling situation acquisition. In order to solve the problems of time-consuming and inefficient manual collection of cotton seedling information, this article explored a new method of extracting seedlings based on unmanned aerial vehicles (UAV) visible light remote sensing images. Firstly, cotton images in the 3-4 leaf stage were captured by the UAV equipped with a high-resolution visible light sensor. Meanwhile, the typical images were selected for experimentation after a series of preprocessing operations, such as correction, stitching, and cropping. The separation of cotton from the background (soil, mulch) was a primary prerequisite for obtaining cotton seedling situation information. The segmentation effect of eight color indexes on cotton image were compared and analyzed and the green-blue difference index (GBDI) color index was selected in this paper to realize the segmentation of cotton and background by combining with the Otsu threshold segmentation method because GBDI component was found to have fewer impurities and more complete segmentation. In order to avoid the influence of weed noise on the follow-up experiment, morphology and grid method for weed noise elimination were adopted, and the results showed that the grid method was more effective than the morphological method in removing weeds. A mapping relationship between morphological characteristics and the number of cotton plants was established to estimate the number of cotton. Because conglutinated cotton was difficult to be segmented by morphological operation, 10 morphological features were extracted as candidate variables to establish SVM plant number estimation model. A total of 3710 sample data were obtained in this experiment, 80% of which were randomly selected for classification modeling, while the remaining 20% were used for testing. Based on the person correlation analysis, 6 features whose correlation coefficient more than 0.7 were selected.The model was applied to the whole image to obtain the number of emerging cotton plants in the study area, and then the seedling emergence rate, canopy coverage and evenness of cotton plant growth were calculated, consecutively. The results showed that the classification accuracy of SVM plant number estimation model reached 97.17%, the statistical error ranged from 0.8% to 4.7%, and the average error was 2.52%. The error of the method decreased with the increase of monitoring area, which indicated that the model had better applicability in larger cotton fields. The cotton emergence rate, canopy coverage and growth uniformity were 64.89%, 7.17%, 10.98% respectively. The method based on the UAV visible light image effectively improved the efficiency of cotton field seedling acquisition, and the research results can provide technical support for subsequent cotton field management and fine plant protection.