联合自动样本生成与样本迁移的关中平原冬小麦识别

Identification of Winter Wheat in Guanzhong Plain Based on Combined Automatic Sample Generation and Sample Migration

摘要:

高质量训练样本对作物遥感识别至关重要，及时、准确地获取冬小麦样本是进行冬小麦遥感识别的基础。但样本点获取往往较为困难，是限制利用卫星遥感影像进行作物分类识别的关键因素。为准确识别关中平原地区多年冬小麦种植区域，本研究基于Sentinel?2卫星遥感影像，构建了适用于该地区冬小麦识别的自动样本生成和样本迁移策略，并且按照市级行政边界将关中平原划分为5个研究区（宝鸡市、咸阳市、西安市、铜川市和渭南市），以2020年作为参考年份，2019年和2021年作为迁移年份，进行冬小麦农田遥感识别方法构建和验证。首先，基于已有的空间分辨率30 m中国冬小麦种植分布数据集开发自动样本生成方法，获取2020年关中平原冬小麦训练样本，并通过咸阳市实测样本验证自动采样的准确性;同时，基于随机森林算法和4个不同生育期（越冬期、返青期、抽穗期、成熟期）遥感影像对冬小麦农田进行分类识别，获得不同生育期冬小麦农田识别结果。然后，基于自动获取的训练样本，利用欧氏距离（Euclidean distance，ED）和光谱角距离（Spectral angle distance，SAD）确定测试区（咸阳市、宝鸡市、西安市）进行样本迁移的最佳生育期和对应的划分阈值。最后，基于优选生育期与阈值利用阈值验证区（渭南市、铜川市）验证识别冬小麦的有效性，并绘制关中平原2019—-2021年冬小麦空间分布图。结果表明：利用冬小麦产品底图进行自动样本生成方法能够得到准确、可靠的冬小麦农田样本，参考年份关中平原冬小麦农田识别总体精度均在93%以上，F1值均高于93%;提取面积与陕西省统计数据相对误差低于18%，与底图数据相对误差低于12%。对关中平原冬小麦样本进行年份迁移时，ED和SAD最优阈值为0.4和0.8，最优生育期为抽穗期，此时5个研究区在2019年和2021年冬小麦农田遥感识别总体精度均大于88%，F1值均高于89%;提取面积与陕西省统计数据（除2019年渭南市）相对误差均低于23%，与底图数据相对误差均低于19%。本文自动样本生成和迁移方法能够较为精准、快速地识别关中平原地区的冬小麦种植区。

Abstract:

High?quality training samples are crucial for crop recognition using remote sensing. The timely and accurate acquisition of winter wheat samples serves as the foundation for such identification. However, obtaining sample points is often challenging, representing a key factor that limits the classification and recognition of crops using satellite remote sensing imagery. To accurately identify multi?year winter wheat planting areas in the Guanzhong Plain, an automatic sample generation and sample migration strategy suitable for winter wheat recognition in this region was constructed based on Sentinel?2 satellite remote sensing imagery. The Guanzhong Plain was divided into five study areas (Baoji, Xianyang, Xi?an, Tongchuan, and Weinan) according to municipal administrative boundaries. Using 2020 as the reference year and 2019 and 2021 as the migration years, a method for remote sensing identification of winter wheat fields was developed and validated. An automatic sample generation method was developed based on an existing 30 m spatial resolution Chinese winter wheat planting distribution dataset to obtain training samples for the Guanzhong Plain in 2020. The accuracy of automatic sampling was verified using measured samples from Xianyang City. Concurrently, the random forest algorithm and remote sensing imagery from four distinct growth stages (overwintering, regreening, heading, and maturity) were employed to classify and identify winter wheat fields, yielding recognition results for each stage. Subsequently, based on the automatically acquired training samples, the Euclidean distance (ED) and spectral angle distance (SAD) were utilized to determine the optimal growth stage and corresponding classification thresholds for sample migration in the test areas (Xianyang, Baoji, and Xi?an). Finally, the effectiveness of identifying winter wheat using the optimized growth stage and thresholds was validated in the threshold verification areas (Weinan and Tongchuan), and spatial distribution maps of winter wheat in the Guanzhong Plain from 2019 to 2021 were generated. The results indicated that the automatic sample generation method utilizing the winter wheat product base map can obtain accurate and reliable winter wheat field samples. For the reference year, the overall accuracy (OA) of winter wheat field recognition in the Guanzhong Plain exceeded 93%, with F1?scores higher than 93%. The relative error between the extracted area and Shaanxi provincial statistical data (except weina) was less than 23%, and compared with the base map data, it was less than 12%. When conducting inter?annual migration of winter wheat samples in the Guanzhong Plain, the optimal thresholds for ED and SAD were determined to be 0.4 and 0.8, respectively, with the heading stage identified as the optimal growth period. Under these conditions, the overall accuracy of remote sensing recognition for winter wheat fields in the five study areas for 2019 and 2021 was greater than 88%, with F1?scores higher than 89%. The relative error between the extracted area and statistical data (except Weina in 2019) was below 23%, and compared with the base map data, it was below 19%. The research result demonstrated that the proposed automatic sample generation and migration method can accurately and rapidly identify winter wheat planting areas in the Guanzhong Plain.