| 摘要: |
| 【目的】面对长三角一体化地区生态高质量修复与智能化监测的更高诉求，开展基于深度学习的水体生境质量评价，旨在探索风景园林数字技术的前沿领域，为长三角一体化地区的水体生态修复与生态绿色发展提供信息化与智能化的技术支撑。【方法】采用基于深度学习的图像分类方法，通过卷积神经网络（convolutional neural networks, CNN）训练，实现大范围、大批量的水体生境卫星图像智能识别、分类与评价。【结果】构建了长三角一体化先行启动区水体生境卫星图像数据集，训练了水体生境质量评价深度学习模型，对研究范围内全域的水体生境进行了高精度、自动化的生境质量评价。【结论】深度学习模型能够长时序、大范围地对水体生境进行质量评价，为水体生境的修复实践提供技术支撑，未来可实现对长三角一体化示范区水体生境质量的跟踪监测。 |
| 关键词: 风景园林 人工智能 水体生境 计算机视觉 深度学习 图像分类 卷积神经网络 长三角 |
| DOI：10.12409/j.fjyl.202302040045 |
| 分类号:TU985 |
| 基金项目:国家自然科学基金“基于多重价值协同的城市绿地空间格局优化机制：以上海大都市圈为例”（编号 52178053）；上海市 2023 年度“科技创新行动计划”软科学研究项目“公园城市目标下上海城市公园生态数字孪生技术创新与精细化管控策略”（编号 23692106200）；自然资源部大都市区国土空间生态修复工程技术创新中心开放性项目“长三角一体化示范区水网复合生境质量智能评价及其修复技术”（编号 CXZX202304） |
|
| Water Habitat Image Classification and Quality Evaluation Based on Deep Learning: A Case Study of the Pilot Zone of the Yangtze River Delta Integration Area |
|
WANG Jieqiong, JIANG Huiqing, WANG Min
|
|
Tongji University
|
| Abstract: |
| [Objective] The Yangtze River Delta (“YRD”) integration area is one of the most typical water network areas in southern China, where river and canal networks are interwoven, and ponds and lakes are widely scattered. Water bodies serve as the lifeline of the water towns in southern China, and it is crucial to improve the water ecosystem services of water bodies therein. With the development of river and lake ecology restoration work, the evaluation of water habitat quality has gradually become a hot topic. Through a literature review, it is found that the shortcomings of existing research are as follows: in terms of research objective, most of existing researches in China focus on rivers in the plain river network area in northern China or the mountainous areas in southern China, while paying less attention to the water network in the Yangtze River Delta; in terms of research content, the focus is on high-precision evaluation of the entire river basin or macro, qualitative research on large-scale areas, and there is still a lack of large-scale, large-sample and refined classification evaluation research; in terms of research method, although there are relatively mature evaluation systems internationally, they are mostly based on field surveys and sampling surveys, which are time-consuming and featured by small sample size, limited evaluation range and poor data tracking, often unable to keep up with the speed of land use and cover change. For the high-density water network in the YRD integration area, existing research methods have obvious technical bottlenecks and fail to meet the practical requirements of ecological restoration planning and design. In response to the higher demand for high-quality ecological restoration and intelligent monitoring in the YRD integration area with a focus on the quality of water habitat, this research proposes that large-scale and large-batch satellite images of water habitat can be intelligently identified, classified and evaluated through the training of convolutional neural networks (CNN) involved in deep learning (DL), aiming to explore the forefront of digital technology in landscape architecture, and provide information and intelligent technical support for the integration of ecological and green development in the YRD area. [Methods] This research proposes an image classification method based on convolutional neural networks (CNN) and utilizes satellite images obtained through network channels in the YRD integration aera as a dataset (including training, testing, and validation sets) for pre-processing. The Urban River Survey (URS) is used as the evaluation index system of image classification for water habit in the YRD integration area, and the dataset is annotated for classification. The deep learning model is trained using the training set, and the accuracy of the model is tested using the testing set. The parameters are continuously adjusted, and the model is evaluated using the validation set. The trained deep learning model can quickly identify satellite images in the YRD integration area and intelligently classify and evaluate water habitat quality. [Results] Water habitat quality can be evaluated from the three aspects of physical habitat, vegetation type and material type, and classified according to the SHQI grading evaluation standard. The results show that the water habitat quality of the YRD integration area should be improved. The number of water habitats with particularly good quality is relatively few, mainly because of high artificialization of water barges, few types of riverway habitats, relatively single vegetation type in the riparian zone, and navigable water transport. Water habitats with relatively good quality often have the following characteristics: a high proportion of natural or naturalized banks, multiple types of habitats such as shallow shoals and streams in riverways, rich vegetation in the riparian zone, various aquatic plants in riverways, few riverway revetments, and materials dominated by mainly biodegradable revetments (such as reeds and wooden stakes) or open cornerstone revetments (such as ripraps and stone cages). In the vast agricultural land, the water habitat quality is relatively poor, generally rated as “Below Average” or “Poor”. The characteristics of these areas include a single type of river habitat, few or no vegetation cover in the riparian zone, and hard revetments in artificially excavated waterways. Due to the highly engineered waterways in urban areas, villages or residential areas, the water habitat quality is mainly rated as “Poor”. [Conclusion] The evaluation of water habitat quality is a crucial aspect of water ecology restoration in the context of booming ecological green integration in the Yangtze River Delta. This research constructs a water habitat quality evaluation index system by selecting evaluation indicators based on image perception, and trains a deep learning model using image classification methods. The application of deep learning models can conduct long-term and large-scale quality evaluation of water habitats, improve work efficiency, expand the spatiotemporal dimension of water habitat quality evaluation, and reveal the changes in water habitat quality. By updating image data, it can track and monitor the water habitat quality in the YRD integration area, and explore the development of digital technology in the field of landscape ecology, and provide technical support for the restoration of water habitat and green development in the YRD area. |
| Key words: landscape architecture artificial intelligence water habitat computer vision deep learning image classification convolutional neural network (CNN) the Yangtze River Delta |
