基于IRCF模型的无人机低空遥感影像田块边界提取方法研究  

IRCF-based unmanned aerial vehicle low altitude remote sensing field boundary extraction method

作　　者：何思涵 胡佳慧 何忠秀 李青涛 王霜[4] HE Si-han;HU Jia-hui;HE Zhong-xiu;LI Qing-tao;WANG Shuang(School of Mechanical Engineering,Xihua University,Chengdu,Sichuan 610039,China;Sichuan Academy of Agricultural Machinery Sciences,Chengdu,Sichuan 610066,China;School of Computer and Software Engineering,Xihua University,Chengdu,Sichuan 610039,China;Institute of Modern Agricultural Equipment,Xihua University,Chengdu,Sichuan 610039,China)

机构地区：[1]西华大学机械工程学院,四川成都610039 [2]四川省农业机械科学研究院,四川成都610066 [3]西华大学计算机与软件工程学院,四川成都610039 [4]西华大学现代农业装备研究院,四川成都610039

出　　处：《南方农业学报》2025年第1期111-123,共13页Journal of Southern Agriculture

基　　金：国家自然科学基金项目(51905447);四川省现代农业装备工程技术研究中心课题(RX2300001944)。

摘　　要：【目的】构建无人机低空遥感影像田块边界提取新方法,为解决丘陵山区小面积田块边界信息提取难题提供技术支持。【方法】利用无人机低空遥感技术获取指定田块的高分辨率影像,制作单个田块图像数据集;在深度学习边缘检测算法RCF模型的基础上,通过减少池化层、增加注意力机制模块,构建准确性能更高的Improve-RCF(IRCF)模型进行田块边界识别,然后对识别后的边界图像进行轮廓检测量化处理,将连续的边界线转换成由多个像素点组成的离散化边界线以便获取像素坐标,再通过高斯—克吕格投影转换为适用于农机作业的平面坐标,并以田块有效面积比和坐标平均偏差2个指标进行田块边界提取准确性验证。【结果】相对于RCF模型,IRCF模型的算法性能得到提升,其检验结果不仅在数据集尺度最佳(ODS)中的F_(1)值领先1.8368%,在图像尺度最佳(OIS)中的F_(1)值也领先2.7969%,且平均精度(AP)提升3.8540%。田块边界线经量化处理后,计算所得的内边界范围内像素总数为21339个,进一步换算求得提取的田块边界线围成面积为1841.32 m^(2),IRCF模型提取率为91.02%。将IRCF模型与Canny测算子边界量化相结合,通过对比分析实际测量值与提取值,发现平面坐标在X轴方向上的平均偏差为0.613 m,在Y轴方向上的平均偏差为0.744 m,均小于0.800 m,说明能提供较准确的边界坐标信息。【结论】基于IRCF模型的无人机低空遥感田块边界提取方法能解决传统边缘检测算法难以识别田块图像数据集中狭窄田块边界的问题,获得的田块边界图像经轮廓检测量化处理后能实现田块边界平面坐标准确获取,实现了丘陵山区小面积田块边界信息的高效率自动提取。【Objective】To develop a new method for extracting field boundaries from low-altitude remote sensing images of unmanned aerial vehicle(UAV),which could provide technical support to address the challenges of extracting boundary information for small fields in hilly and mountainous areas.【Method】Utilized low-altitude remote sensing technology from UAV to acquire high-resolution images of specified fields,creating a single-field image dataset.Based on the RCF model,a deep learning edge detection algorithm,an improved RCF(IRCF)model with higher accuracy and performance was constructed by reducing pooling layers and incorporating an attention mechanism module for field boundary recognition.Subsequently,the recognized boundary images treated by contour detection and quantization processing,converting continuous boundary lines into discrete boundary lines composed of multiple pixel points to obtain pixel coordinates.These coordinates were then transformed into plane coordinates suitable for agricultural machinery operations using the Gauss-Krüger projection.The accuracy of field boundary extraction was validated using 2 indicators:the effective field area rate of field and the average coordinate deviation.【Result】Compared to the RCF model,the IRCF model demonstrates improved algorithmic performance.Its test results showed a 1.8368%lead in the F_(1) score at the optimal dataset scale(ODS)and a 2.7969%lead in the F_(1) score at the optimal image scale(OIS),with an average precision(AP)increase of 3.8540%.After quantization processing of the field boundary lines,the total number of pixel points within the inner boundary area was calculated to be 21339,which further converted to an extracted field boundary area of 1841.32 m^(2),with an extraction rate of 91.02%for the IRCF model.By combining the IRCF model with Canny operator boundary quantization and comparing the actual measured values with the extracted values,it was found that the average deviation of plane coordinates in the X-axis direction was 0.613 m,and

关 键 词：田块边界 IRCF模型 边缘检测算法 无人机低空遥感影像 坐标获取 

分 类 号：S127[农业科学—农业基础科学]