Image-based Automatic Observation Technology For Critical Growth Stages Of Rice

With the growth of world’s population quantity, food problems have become a pressing problem nowadays. China is the country with the largest population in the world, but the per capita arable land is far lower than the world average. Therefore, food security problems have brought our country serious challenges. As rice is one of the important food crops, its production has been drawn much attention in our country. The yield of rice can be affected by climate conditions and agricultural production management. The observation of rice growth periods not only can guide the implementation of the production management according to the arrival of growth periods, but also help study the necessary of agriculture meteorological conditions when the growth periods arrival. Indeed the observation of rice growth periods can ensure the high and stable yield of rice. At the same time, rice growth periods observation plays an important role in modern agriculture informationization management and phenological observation.At present, the observation of rice growth periods still depends on the manual observation by farmers. However, manual observation has many disadvantages, such as time-consuming, laborious, long observation interval and large subjectivity of the observer, etc. In agriculture production, original way of manual observation gradually cannot satisfy the needs of modern agriculture management and the need of phenological study. In this dissertation, growth state of rice in paddy field is automatically observed through an automatic observation device. With the collected rice image and observation experience of farmer, computer vision technology is used into the study of in the stages of rice transplanting, tillering and heading.Accurate extraction of rice from image is an important process in the automatic detection of rice growth periods, which can affect the accuracy of the subsequent detection results. In order to best solve the segmentation problem of rice image taken under complex environment such as serious illumination changes, the strong light reflection of water surface and the shadow, in this dissertation we studies the previous crop segmentation algorithms, and the research achievements of skin color segmentation, and proposes a rice segmentation method based on morphological modeling at last. This method establishs the color model of rice in images. It overcomes the disadvantages of the previous threshold, single gaussian model and Bayesian theory based crop segmentation algorithms to a certain extent. Through the study of images under different illumination conditions, this method can well adapt to the natural environment of illumination changes. With the proposed segmentation method, well segmentation results of rice can been achieved. Moreover, the proposed method can laid the foundation for the subsequent research.Rice transplanting represents the beginning of the rice growth in paddy field. It is an turning point in the process of rice growth. For the automatic detection of rice transplanting stage, this article analysed the images taken in the rice transplanting, and then selected two kinds of image features which are the jump of the rice coverage value and rice distribution uniformity to represent the transplanting of rice. Using the above two kinds of image features, this method can automatically detect the finish of rice transplanting in the observation region. The experimental results show that the automatic detection results of this method are consistent with manual observation.Rice tillering stage determines the number of rice headings in heading stage. It’s an important growth period which can affect the yield of rice. Moreover, it can guide the farming activities such as fertilizer and insecticide, etc. and ensure the high yield of rice. Though the analysis of physiological characteristics of rice tillering and the obtained image, this dissertation proposes a multi-feature description and random forest classification based automatic detection method of rice tillering stage. This method combines rice coverage, junction points of skeleton image and Harris corner points as the image features for the rice tillering. Finally, this method adopts the random forest classification to give the automatic detection result of tillering stage. Experimental results verified the accuracy of this algorithm through comparison of the automatic detection results and the record of manual observation.Rice heading stage determines the value of rice grain filling rate. It is another important growth period which can affect the rice yield. Indeed, rice heading stage observation is always an important aspects of the observation. This dissertation presents an automatic detection method of rice tillering stage, which can automatically detects the heading stage of rice based on the ear color feature, gradient histogram detection and convolution neural network. The method adopts convolutional neural network to automatically learn the underlying image features of rice images. Afterwards, those learned underlying features are used to complete the automatic detection of rice heading stage. This method overcomes the rice leaf color change and the influence of the low resolution of image to a certain extent. Experiment showed the feasibility of the proposed heading stage detection method through comparison of the automatic detection results and the record of manual observation.Between2011to2013, we collected a total of12rice image sequences through the automatic observation system of rice growth status, which includes6early rice image sequences and equal number of late rice image sequences. In this dissertation, the effectiveness of the proposed methods are verified by using the above12image sequences. The differences between the results of automatic detection methods and the manual observation records are almost within three days, which meet the actual needs of the agricultural meteorological observation. The proposed automatic detection methods can be used to simplify or replace the tedious manual observation.