Research On Image-based Technology For Automated Dew And Frost Detection

As the weather features that happen near the land surface, dew and frost have significant impacts on the thermal balance of atmosphere, the water and the energy balance of eco-logical systems, and the nutrient cycling of crops. Both dew and frost can be obtained by meteorological observations, especially, by surface weather observations, which provide the fundamental data to meteorology researches and meteorological services such as weather forecasting. However, dew and frost observations still mainly rely on manual labor, which can not meet the requirements of current operational observations.In manual observations, the visual information of various weather features is the most convincing evidence for obtaining the present weather features. From the perspective of computer vision, the visual information can be represented by image features which can be further recognized in terms of an automated way. Therefore, this dissertation carries out several researches on automatically observing dew and frost based on computer vision techniques and machine learning theories.Firstly, we design a new artificial contact surface for dew and frost deposition, which facilitates the image-based observations of both weather features. Furthermore, the repre-sentativeness of the contact surface to natural vegetation is also discussed in both theoretical and experimental manner. To precisely locate the contact surface in the image, we propose a novel saliency detection method for extracting the contact surface from images. Our saliency detection method is also capable of detecting attractive regions in the unspecified natural im-ages, and the effectiveness and the robustness of our model are validated on5commonly used benchmarks.Secondly, we propose a new image-based method for dew observation. By analyzing the optical changes of the contact surface caused by dew deposition, several low-level visual features are explored to describe the optical properties. Then, the change regularity of these features are further used to reveal the process of dew formation. Thirdly, based on the observations of image series obtained during a frost formation pe-riod, we propose a new method for observing frost. We first use a correlation method to measure the similarities between the images. These similarities are then fitted to a numeri-cal function that describes the frost formation on the contact surface. Additionally, a SVM classifier is learned to classify the texture information of frost.The types of natural vegetation planted in the observation field are diverse, and even a certain type of vegetation may varies in different seasons, which potentially causes a de-crease in the performance of frost observation as the classifier is ineffective to recognize the unlearned patterns. We thereby propose another method for observing frost via saliency la-beling and on-line learning. Firstly, we introduce a novel top-down and task-driven saliency detection model, which takes the color information of the typical frost images as the prior. For each image, a manifold learning procedure is adopted to rank the saliency of superpixels that are extracted from the image. The salient value of a superpixel can also be regarded as the probability of the superpixel that refer to the region with frost deposited, and we thereby use a simple adoptive one dimensional thresholding method to separate all the superpixels into two categories:frost regions and dry regions. The saliency detection facilitates our method in two ways:on the one hand, the dry regions that are annotated by the saliency model and the manually annotated frost regions can be used to train or update a SVM classifier before frost detection in a day. On the other hand, one can merely concern on the frost regions that are annotated by the saliency model.Finally, the applications of the proposed dew and frost detection algorithms in daily weather observations are discussed. A full functional system is developed to facilitate the observation tasks in standard meteorological stations. Besides the observations of dew and frost, this system also integrates the observations of precipitation, wire icing, freezing and snow depth measurement using the image-based methods. The system has been applied in several meteorological stations in China, which shows the practical value of our researches.