Illumination-invariant Segmentation And Growth Stages Prediction For Crop Research

Precision agriculture is the trend of the world's agricultural development. It is committed to guide agricultural production, to provide information for agricultural activities and to monitor and supervise the crop growth stages. The computer vision and pattern recognition technology have been applied in precision agriculture, which has already achieved some applications like disease surveillance, weed identification and robot guidance on the field. However, these are still difficulties and challenges in precision agriculture, including crop extraction, biometric feature obtainment and the crop growth stages supervision.In the early growth of crop, seedling leaf is small, the background is complex and projected shadows on the land are greenish, which brings many difficulties to extract the crop. Therefore, this thesis presents a seedlings extraction method under a complex background. Utilizing the spatial distribution of different color spaces under different lighting conditions, the segmentation algorithm firstly extracts 15 color channel features. It uses random forest to build the decision trees and analyzes effect of different color spaces for crop extraction. Then performances of crop extraction under shadow from different color spaces have been analyzed. And a crop extraction approach based on support vector machine(SVM) algorithm has also been proposed. This method achieves a stabilized crop segmentation performance with the ability to extract the seeding from complex background. Even if there is a shadow in the image, the support vector machine(SVM) one can achieve a better crop extraction performance, compared with other approaches.As the crop enters into the reproductive growth stages, leaf area began to grow vigorously. Light start to have a great impact on the leaf at this time. When light is strong, there is a large area of specularity on the leaf, which again brings difficulties for crop extraction. From our human vision, there exists a color gradient area between the high light area and its neighbors. According to this motivation, the thesis proposes a specularity-invariant crop extraction approach based on the Markov model. The algorithm uses superpixels and statistical data to establish the potential function. And it uses loopy belief propagation algorithm to obtain optimal solution. Comparison experimental data shows that the algorithm is resistance to high light and achieves the highest mean and the lowest variance in the total of 158 testing images which includes different lighting conditions and crop categories.In the traditional agriculture observing guidance, it requires means of different human senses to measure the crop physically, which is difficult to obtain the absolute measurement information from the image. Using computer vision technology, the thesis presents eight categories of biomass characteristics based on image to reflect the morphological characteristics at different crop growing stages. Then this thesis proposes the monitoring and prediction algorithm for crop growth stages based on the proposed biomass feature,where the extreme learning machine has been utilized to build the prediction model. The experiment demonstrates that the result of the predictions model and the manual observations have little difference.