综合样本、特征和分类器多要素的作物分类对比研究

Comparative Analysis of Crop Classification by Integrating Multiple Factors of Samples, Features, and Classifiers

关键词:  遥感  作物分类  多要素  对比研究

Key words:  remote sensing  crop classification  multiple factors  comparative analysis

摘要:

【目的】利用遥感技术能够大范围快速提取作物的空间分布信息,其精度是由样本、特征、分类器等多要素共同决定的。当前研究主要关注单要素和双因素的对比分析,缺少多要素的作物分类对比研究。【方法】文章基于“MAP杯”数智农业大赛作物识别项目,收集了2021年湖北枝江区域的13个作物分类结果以及2022年山东桓台区域和湖北枝江区域的14个作物分类结果,通过构建的分类要素体系对比了样本、特征、分类器要素及其组合的精度差异,并按不同的作物类型进行了一致性评价。【结果】研究结果表明：(1)2021年和2022年使用最多的多要素组合均为目视解译样本-光谱 植被指数特征-传统机器学习分类器,平均精度最高的分类要素分别为实地采集样本、高层特征以及传统机器学习分类器；(2)玉米、小麦等主粮作物的空间分布一致性较高,油菜等经济作物以及复合种植的作物类型的空间分布一致性低,是区域作物制图的难点。【结论】时空迁移样本-高层特征-深度学习分类器的要素组合未来仍存在较大发展潜力。文章能为高效率高精度地解决作物识别任务提供参考和依据。

Abstract:

[Purpose] Remote sensing can quickly obtain spatial distribution information of crops on a large scale, and its accuracy is influenced by multiple factors such as samples, features, and classifiers. The current study mainly focused on the comparative analysis of one or two factors but lacks the multi-factor comparisons for crop classification. [Method] Based on the crop identification project of the MAP competition, this paper collected 13 crop classification results of Zhijiang in 2021 and 14 results of Huantai and Zhijiang in 2022. The overall accuracy of different classification factors and combinations was compared and analyzed based on the system of classification factors, and the consistency of spatial distribution was evaluated for different crops. [Result] The results showed that the most dominant multi-factor combination in both 2021 and 2022 was visual interpretation samples-spectral vegetation index features-traditional machine learning classifiers. The field samples, high-level features, and traditional machine learning classifiers showed the highest average overall accuracy under different classification factors, respectively. The spatial distribution of staple crops such as corn and wheat had high consistency, while cash crops such as rape and compound planting crops varied widely and had low consistency. [Conclusion] The factor combination of spatio-temporally transferred samples-high level features-deep learning classifiers still has a large potential for future crop classification. This paper could provide reference for solving the crop classification task with high efficiency and high accuracy.