复杂场景下小农经营区地块级苹果园模块化制图方法框架

Modular Mapping Framework of Parcel-Level Apple Orchards in Smallholder Agriculture Areas under Complex Scenes

实现复杂场景下地块级苹果园的精准制图,是中国苹果产业精细化管理面临的迫切需求。然而,传统的地块级分类制图框架在处理小农经营区内的大量细碎地块时,可靠性大幅度降低。本文提出一种适用于复杂场景下小农经营区的地块级苹果园模块化制图方法框架。①基于模拟人类对目标场景视觉感知的分层策略,从深秋季的单幅超高空间分辨率(Very High Resolution,VHR)影像中提取冗余的候选地块;②利用非对称瓶颈网络(Depth-wise Asymmetric Bottleneck Network,DABNet)模型与长短期记忆(Long Short-Term Memory,LSTM)模型,分别从VHR影像与时间序列影像中提取苹果园的空间特征像素与时序特征像素。然后,构建元特征描述特征像素在候选地块中的分布情况,与地块的内在特征共同组成苹果园地块的分类特征;③使用随机森林(Random Forest,RF)将候选地块分类为苹果园地块和非苹果园地块。以山东省栖霞市西城镇为研究区,从43238个候选地块中提取出30292个苹果园地块,分类总体精度达到92.7%。利用RF算法自带的平均精度减少指标(Mean Decrease in Accuracy,MDA)对17种分类子特征进行特征重要性分析,证明本文提出的元特征比传统人工设计特征具有更强的信息抽象与特征表达能力。该框架成功实现场景复杂的小农经营区地块级苹果园制图,可推进精准果园农业的发展。

Accurate parcel-level mapping of apple orchards in complex scenes is an urgent need for refined management of China's apple industry.However,the traditional parcel-level classification mapping framework is substantially less reliable when dealing with a large number of fragment and small parcels in smallholder management areas.In this paper,a modular parcel-level mapping framework of apple orchard was proposed,which was suitable for smallholder management areas under complex scenes.First,the hierarchical strategy which simulates human visual perception was used to extract redundant candidate parcels from a single Very High spatial Resolution(VHR)image in deep fall.Second,we used the Depth-wise Asymmetric Bottleneck Network(DABNet)to extract spatial feature pixels from a VHR image and used the Long Short Term Memory(LSTM)to extract time series feature pixels from optical time series images.Then,meta-features were constructed to describe the distribution of feature pixels extraction results within the parcel,which together with the parcel intrinsic features formed the classification feature of the apple orchard parcels.Third,the Random Forest(RF)algorithm was used to classify the candidate parcels into apple orchard parcels and non-apple orchard parcels.An experiment was carried out in the Xicheng Town,southwest of Qixia City,Shandong Province,China.30292 apple orchard parcels were extracted from 43238 candidate parcels,and the Overall Accuracy(OA)turned out to be 92.7%.Using the Mean Decrease in Accuracy(MDA)of RF algorithm to analyze the importance of 17 classification features,it was proved that the proposed meta-features had stronger spatial information description ability and feature expression ability than traditional features.This framework successfully realizes parcel-level mapping of apple orchards in smallholder agriculture areas under complex scenes,which can promote the development of precision orchard agriculture.