Research On Vision-based Automatic Observation Technology For Growth Status Of Cotton

Smart agriculture is an inevitable trend in the development of modern agriculture.Agricultural mechanization and information intelligence will play key roles in the progress.However,current development speed of information intellectualization is slow,with low level of automation.Crop growth information is an important scientific guidance basis for agricultural operations.Nowadays,manual observation records still are the primary means to acquire crop growth information,which are time-consuming and subjective.Therefore,there is an urgent need to adopt automated observation methods to overcome the“incomplete”and“uncertain”problems caused by manual observations.It can provide reliable intelligent information for mechanized agricultural operations,and offer opportunities for production surging and labor liberation.This thesis takes cotton as the research object,which is a vital source of important economic crop and natural fiber raw material in China.On basis of the collected in-field cotton images acquired by ground-based observation device and the observation experience of agricultural meteorological observers,this thesis explores how to identify cotton growth status with computer vision technologies,aiming to automatically detect the key growth stages of cotton.Since the field management and agricultural operations in the boll splitting stage and the boll opening stage directly affect the yield and quality of cotton,the automatic detection methods in these two stages will be the focus of this study.The major contribution of this thesis is concluded as following:To address the problem of varying poses,unpredictable scales brought by cotton bolls,this thesis has proposed a new method for cotton boll splitting stage detection,based on multi-scale feature coding and morphological transformation.According to the observation definition and image characteristics,the method is divided into two sub-tasks: cotton boll detection and cotton boll splitting state detection.Cotton boll detection can be implemented by a linear support vector machine classifier.The image representation vector which should be sent to the classifier is generated by locality-constrained linear coding and spatial pyramid matching model.Thus,the image representation vectors with spatial and semantic information can improve the detection accuracy.As for boll splitting stage detection,extracting cracks in the detected image block of cotton boll with morphological transformation technologies will realize the automatic detection of cotton splitting stage.The experimental results compared with the manual observation records verify the accuracy and effectiveness of proposed boll splitting stage detection method.To tackle the problems of weak robustness and poor applicability in traditional cotton segmentation methods,this thesis has reported on a novel boll opening stage detection method via region-based semantic segmentation,including unsupervised region generation,region feature extraction and supervised semantic labeling prediction with random forest.This method is the first for solving cotton segmentation issue at the regional level.Simple linear iterative clustering and density-based spatial clustering method with metric learning are employed to generate regions,with superiority in edge-preserving and density contrast distribution.The experimental results on the constructed dataset show that the proposed method can adapt well to the complex farmland environment and obtain better cotton segmentation results.The experimental results compared with the manual observation records also verify the accuracy and effectiveness of proposed boll opening stage detection method.To solve the problems of limited discriminative power and poor applicability to changing weather brought by hand-crafted features,this thesis has proposed a boll opening stage detection method based on fully convolutional network.Different from hand-crafted features,the deep convolutional networks guarantee powerful feature description capability,because they can extract the cotton signature through a data-driven way,achieving low-level and highlevel semantic information,simultaneously.Thus,this proposed method first introduce deep learning to in-field cotton segmentation.Adopting coarse-to-fine strategy on initial segmentation image generated by fully convolutional neural network,cotton segmentation image can be achieved by contextual information refinement.The experiments demonstrate that the proposed method outperforms other state-of-the-art approaches on constructed dataset,either in different common scenarios or single/multiple plants.In addition,the proposed method can improve the detection accuracy of boll opening stage as well.In order to realize the intelligence of key growth stage information acquisition,this thesis has designed a set of agriculture intelligence aid system,which is called integrated software system for automatic cotton growth stage observation.The system can record,observe and analyze cotton growth status continuously,by means of integrating seven proposed key growth stages detection methods of cotton.The experimental results demonstrate that the system can detect the growth stages accurately,with comparable performance to manual observation.This validates studies covered in this thesis are of great application promotion and practical use value.