摘要
快速精准地掌握作物种植模式信息对于农业产能评估、区域种植结构平衡和国家粮食安全保障具有重要的战略意义。然而,目前尚缺乏高精度、大范围、涵盖复杂种植模式的空间数据集。因此,本研究以中国南方重要的产粮地之一江汉平原为例,基于GEE平台和Sentinel-NDVI数据,构建时序物候特征集,通过探究半自动提取大量样本的方法,对比多粒度级联森林(multi-grained cascade Forest,gcForest)与深度神经网络(Deep Neural Network,DNN)两种深度模型分类精度,对江汉平原5种作物类型的6种种植模式进行精细识别与制图研究。结果表明:1)江汉平原主要包括单季稻、莲藕两种单季种植作物;小麦-水稻/棉花、油菜-水稻/棉花等4种年内复种模式;2)基于半自动采样点(4 000个)的gc Forest模型总体精度最高,可达到87.25%;两种模型基于4000个采样点的分类精度相较基于400个实地采样点分别有8.08和5.5个百分点的提升,该半自动提取样本点的途径可有效提高分类精度。该研究证明,基于物候的Sentinel-NDVI数据在复杂农业景观区域的作物种植模式制图中有较大潜力。
Rapid and accurate extraction of cropping patterns is of great significance for regional resource capacity evaluation, green and sustainable agricultural development and national food security. However, there are few spatial datasets with the precision, breadth of coverage, or sufficient information as required in the mapping of complex crop types and rotation patterns. Furthermore, the majority of existing research on crop extraction focuses on extracting crops either from high-resolution images in small areas or from low-to-medium resolution images in large areas, thus missing complex and dynamic cropping patterns. To make up this shortage, a phenology-based crop type and the mapping technique of cropping patterns were proposed based on the GEE platform and Sentinel-2 time series imagery for high-accuracy crop mapping of large areas. The Jianghan Plain, an important grain-producing region in south China, was studied. The method for semi-automatic extraction of a large number of samples was explored, and the classification accuracy of multi-grained cascade Forest(gc Forest) and that of the Deep Neural Network(DNN) were compared to identify and map the types and cropping patterns of rice, wheat, rapeseed, cotton and lotus root in Jianghan Plain. The results showed that: 1) the main crops in the Jianghan Plain include rice growing in a single season, lotus root, wheat, rapeseed, and cotton. The findings also highlight the major single-season cropping structures, including single rice and lotus root, and the major rotation patterns, consisting of wheat-rice, wheat-cotton, rapeseed-rice, and rapeseed-cotton;2) the formation and distribution of different crop types and crop rotation patterns are driven by multiple factors, such as climate, topography, socioeconomics, and farmers’ subjective wishes: rice is the most widely planted crop in the Jianghan Plain, especially single rice. In recent years, as the urbanization is intensified and the labor force in agricultural planting is aging, many farmers choose to plant single-season rice due to the lack of labor. Therefore, the planting area of single-season rice dominates and expands;wheat is mainly distributed in the western and northern regions of the plain, and its planting areas are relatively concentrated. Areas sown to wheat in the northern part of the plain mainly carry out wheat-soybean and wheat-rice-based crop rotation patterns, while wheat-cotton and wheat-rice patterns are more frequently seen in the western areas along the river. With the promotion of the national policy on farming with machine, areas sown to wheat in Jianghan Plain are expanding year by year;rapeseed is mainly distributed in the middle of the Plain, mostly in the modes of rapeseed-cotton and rapeseed-rice rotation, for rapeseed and cotton have a large demand for water at each growth stage, and are more suitable for planting in areas with moist soil;3) the gc Forest model based on 4 000 semi-automatic sampling points has the highest overall accuracy, which can reach 87.25%. The classification accuracy of the two models based on 4 000 sampling points is 8.08 and 5.5 percentage points higher than that based on 400 field sampling points respectively. This semi-automatic extraction approach of sampling points can effectively improve the classification accuracy. Besides, whether based on 400 or 4 000 sample points, the classification accuracy of the gc Forest model is higher than that of the DNN, indicating that gcForest has an edge when extracting planting patterns in complex agricultural landscapes in south China. In conclusion, the results demonstrate that the phenology-based open-source Sentinel-2 sequential data can effectively support the mapping of planting patterns in large and complex agricultural areas. Thus, the mapping of crop planting patterns in south China presented in this study provides a scientific basis for formulating policies related to crop rotation and sustainable agricultural development.
作者
张紫荆
华丽
郑萱
李嘉麟
Zhang Zijing;Hua Li;Zheng Xuan;Li Jialin(School of Resource Environment,Huazhong Agricultural University,Wuhan 430000,China;School of Urban Planning and Design,Shenzhen Graduate School,Peking University,Shenzhen 518000,China)
出处
《农业工程学报》
EI
CAS
CSCD
北大核心
2022年第1期196-202,共7页
Transactions of the Chinese Society of Agricultural Engineering
基金
国家青年科学基金项目(41601280)。
关键词
模型
时序数据
半自动提取
种植模式
models
time series data
depth model
semi-automatic extraction
planting pattern
