【作者】 丁伟；

【导师】 黄河；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2021， 硕士

【摘要】 我国是农业大国,但是由于地理环境、气候等因素导致自然灾害不可避免。农业保险基于风险转移机制保障农业灾后重建、恢复农业生产,是国家扶持农业发展的重要方式之一。随着遥感技术的发展,越来越多的遥感图像被用于农业保险业务,特别是承保业务。而承保业务核心是要准确地实现地块级别的耕地提取与作物识别。因此,本文基于深度学习展开对这两部分问题的研究,具体研究内容如下:在耕地提取方面,当前主要的方法是将耕地提取建模成自下而上的边缘提取任务,需要再进行一系列边缘闭合、区域分割等后处理算法才能得到地块对象。该方法不仅处理效率低,而且容易产生累积误差。本文将耕地提取建模成自上而下的实例分割任务,从而端到端地提取出地块对象。为了提高基于轮廓建模的实例分割模型精度,本文提出Polar R-CNN模型,将极坐标建模思想引入到二阶段实例分割模型中,在传统的二阶段分割模型基础上加入极坐标分支,同时监督实例轮廓与掩膜,端到端地输出有限个轮廓点,从而获取到地块实例。在作物识别方面,由于不同时期遥感影像的作物特征不同,导致数据分布产生变化,从而严重影响监督模型识别精度。本文使用领域自适应模型应对数据分布变化问题,实现跨时域作物识别。本文在CDAN模型基础上提出一种基于可学习样本权重的CDAN-SWN模型。一方面,针对遥感影像、作物本身特点,通过加宽骨干网络,使用多时相为下游迁移任务提供丰富的特征;另一方面,为了解决困难样本给模型带来的负迁移问题,使用可学习的样本加权网络代替原模型直接计算熵的方式,来更好地度量样本的可迁移性。最后,为了验证本文提出的耕地提取与作物识别模型的有效性,本文以江苏省仪征市为试验区,以谷歌影像和多期哨兵二号卫星影像为数据源,进行实地采样和标注。实验表明:在二阶段实例分割模型中引入轮廓极坐标分支能够有效地提高基于轮廓建模的模型精度,耕地提取的AP指标达到61.5%;使用本文提出的改进CDAN-SWN模型在多种跨时域水稻识别场景下的指标均有较大提升,最终识别精度可达97%。更多还原

【Abstract】 China is a big agricultural country,but natural disasters are inevitable due to geographical environment,climate and other factors.Based on risk transfer mechanism-agricultural insurance guarantees post-disaster reconstruction and recovery of agricultural production,which is one of the important ways to support agricultural development.With the development of remote sensing technology,more and more remote sensing images are used to carry out agricultural insurance business,especially under-writing business.The core of underwriting business is to accurately realize the extraction of cultivated land and crop identification at the plot level.In this thesis,the problems of these two parts are studied respectively.In terms of cultivated land extraction,the current main method is to model cultivated land extraction as a bottom-up edge extraction task,which requires a series of post-processing algorithms such as edge closure and region segmentation to obtain plot objects.This method is not only low in processing efficiency,but also prone to accumulative errors.In this thesis,the extraction of cultivated land is modeled as a top-down instance segmentation task,so as to extract the land object from end to end.In order to improve the accuracy of the single-stage contour-based instance segmentation model,this thesis introduces the idea of polar coordinate modeling into the two-stage instance segmentation model,the polar coordinate branch is added,and the instance contour and mask are supervised at the same time.Finally,an instance of the plot is obtained through a limited number of contour points.In terms of crop recognition,the crop features of remote sensing images in different periods are different,which leads to changes in the distribution of data features and seriously affects the recognition accuracy of the supervision model.In this thesis,a domain adaptive model is used to deal with this problem and realize cross-temporal crop recognition.Based on CDAN model,a crop recognition method based on the learnable sample weight CDAN model is proposed.On the one hand,according to the characteristics of remote sensing images and crops,the model provides rich features for downstream migration tasks by widening the backbone network and using multi-time images;at the same time,in order to solve the problem of difficult transfer caused by hard samples in training,a learnable sample weighted network is used instead of calculating entropy directly to measure the transferability of samples better.In order to verify the validity of the cultivated land extraction and crop identification model proposed in this thesis,the thesis uses Yizheng City,Jiangsu Province as the experimental area,and uses google satellite images and multi-phase Sentinel-2 satellite images as data sources to conduct on-site sampling and manual labeling.Experiments show that the introduction of the contour polar coordinate branch in the two-stage instance segmentation model can effectively improve the accuracy of the contour modeling-based model,and the AP index of the plot extraction reaches 61.5%;Using the improved CDAN-SWN model proposed in this thesis,the metrics in a variety of rice identification data transfer scenarios have been greatly improved,and the final identification accuracy can reach 97%.更多还原