| 基于深度卷积网络和分水岭分割的耕地地块识别方法 |
| |
| 引用本文： | 朱昱,潘耀忠,张杜娟.基于深度卷积网络和分水岭分割的耕地地块识别方法[J].地球信息科学,2022,24(12):2389-2403. |
| |
| 作者姓名： | 朱昱 潘耀忠 张杜娟 |
| |
| 作者单位： | 1.北京师范大学地理科学学部，北京 1008752.北京师范大学遥感科学国家重点实验室，北京 100875 |
| |
| 基金项目： | 国家自然科学基金项目(42192581) |
| |
| 摘 要： | 耕地地块作为精准农业的重要支撑，现有地块边界大多依靠人工勾绘。随着遥感技术的发展，基于遥感影像自动提取耕地地块成为研究主要方向，其中基于深度学习的方法能够克服传统检测方法难以适应复杂场景的局限而被广泛使用，但现有检测方法仍存在问题，基于深度卷积模型直接识别耕地区域会丢失内部边界、而基于边缘检测模型识别耕地边界时则会同时得到大量无关边界；此外，现有的基于阈值提取地块的策略所提取的地块不够规整，存在内陷的问题。针对上述问题，本研究提出一种基于深度卷积网络和分水岭分割的耕地地块提取方法，从信息检测和地块提取两方面进行改进：① 将耕地边界视作一种地物类别，在深度卷积网络中进行类别概率检测，帮助实现对耕地边界的语义识别。② 基于改进后的D-LinknetXt网络进行检测，其网络架构适合于对耕地边界这类线性目标的提取，同时更换原始D-Linknet网络的残差单元，帮助提高了网络的特征提取能力。③ 基于分水岭分割对耕地地块进行提取，利用了区域分割方法获取边界的封闭性，并且这种以区域为单元进行分割并合并的方式，解决了原有方法在像元尺度上基于阈值提取所遇到的提取地块存在内陷的问题，使地块更规整准确。在一景高分二号影像上进行检验，并设计两类精度指标，从全局边界精度和地块几何精度两方面进行方法性能评估。实验结果表明，本方法的精度优于3种对比方法，在代表边界精度的F1分数上达到了0.933，地块几何精度为0.829。本研究所提出的方法能够更加准确的对耕地地块进行提取，并适用于实际应用中。
|
| 关 键 词： | 耕地地块 地块边界 深度卷积网络 分水岭分割 高分辨率影像 遥感 |
| 收稿时间： | 2022-06-27 |
An Agriculture Parcel Identification Method based on Convolutional Neural Network and Watershed Segmentation |
| |
| ZHU Yu,PAN Yaozhong,ZHANG Dujuan.An Agriculture Parcel Identification Method based on Convolutional Neural Network and Watershed Segmentation[J].Geo-information Science,2022,24(12):2389-2403. |
| |
| Authors: | ZHU Yu PAN Yaozhong ZHANG Dujuan |
| |
| Institution: | 1. Faculty of Geographical Science, Beijing Normal University, Beijing 100875, China2. State Key Laboratory of Remote Sensing Science, Beijing Normal University, Beijing 100875, China |
| |
| Abstract: | As an important support for precision agriculture, the acquisition of agricultural parcels mostly relies on extensive manual delineation. With the development of remote sensing technology, automatic extraction of cropland parcels based on remote sensing images has become the main research direction. In particular, the method based on deep learning has been widely used and can overcome the limitation of traditional detection methods that are difficult to adapt to complex scenes. However, there are still problems in existing deep learning methods, e.g., direct identification of cultivated areas based on deep learning models results in loss of internal boundaries, and identification of cultivated areas based on edge detection models results in inclusion of irrelevant boundaries. In addition, the extracted parcels using existing strategies based on thresholds are usually irregular and has the problem of inward trapping. To address the above problems, this study proposes a method of extracting cropland parcels based on deep convolutional neural network and watershed segmentation, which improves both information detection and parcel extraction. First, we treat cropland boundaries as a type of feature category and perform category probability detection in deep convolutional neural network to help achieve semantic recognition of cropland boundaries. Second, we perform parcel boundary detection based on the modified D-Linknet network, whose architecture is suitable for extraction of linear targets such as cropland boundaries, and the replacement of residual blocks can help improve the feature extraction capability of the network. Third, we extract agricultural parcels based on watershed segmentation, using the region segmentation method to obtain the closure of boundaries. The way of segmenting and merging parcels using region as a unit overcomes the limitation of parcel inversion in traditional methods based on thresholds at the pixel level. The performance of our proposed method is evaluated by two types of accuracy indicators: global boundary accuracy and parcel geometry accuracy. The experimental results show that the accuracy of our method is better than three comparison methods, and the F1 score representing the boundary accuracy is 0.933 and the parcel geometry accuracy is 0.829. Our proposed method can extract the agriculture parcels more accurately and is suitable for future practical applications. |
| |
| Keywords: | cropland parcel extraction parcel boundary convolutional neural network watershed segmentation high resolution images remote sensing |
|
| 点击此处可从《地球信息科学》浏览原始摘要信息 |
| 点击此处可从《地球信息科学》下载免费的PDF全文 |
|
