基于时序Sentinel-2A影像的玉米秸秆覆盖区智能识别研究

Intelligent Recognition of Corn Residue Cover Area by Time-Series Sentinel-2A Images

秸秆还田是减少土壤侵蚀、增加土壤有机碳的重要措施,对黑土地保护具有重要意义。区域范围内玉米秸秆覆盖区的准确、快速识别,对监测保护性耕作实施、农业补贴政策的制定具有重要作用。以实施保护性耕作的典型区吉林省四平市为研究区,基于GEE(google earth engine)云平台,结合2020年5月-11月的Sentinel-2时序遥感影像,依据玉米生长季和收获后的秸秆状态构建光谱特征和指数特征,指数特征包括归一化差值植被指数(NDVI)和归一化差值秸秆指数(NDRI)。为避免数据冗余,对时序特征值按大小排序,同时利用分位法以0%,25%,50%,75%,100%分位选取分位(QT)特征,进而构建数据集。应用参数优化后的随机森林方法对按照7∶3划分的样本集进行训练和验证,然后对数据集分类,结合连通域标定法去除分类过程中产生的细小连通域,进一步优化全局结果。通过Kappa和整体精度(OA)定量和定性评价,实验结果表明:(1)基于不同特征集组成数据集的分类模型(M1/M2/M3/M4/M5)定量评价结果均优于90%,其中所设计数据集的分类模型M5效果最好,Kappa和OA分别为97.41%和97.91%,相比于未加入QT特征集的分类模型M2的Kappa和OA分别提升4.52%和3.64%,同时M5识别结果可以有效保留边缘细节信息;(2)针对不同时间尺度的QT特征集,利用5月-11月时序遥感影像的QT特征集分类模型M5_6/M5可以极大地抑制其他作物秸秆的影响,相比仅利用11月时序影像QT特征的M5_1模型分类结果的Kappa和OA分别提升了3.9%和3.12%;(3)基于M5模型,结合连通域标定法的分类模型M6的Kappa和OA分别为96.76%和97.36%,仅次于M5模型识别结果,模型M6在保证较高精度的同时避免了细碎图斑,优化了分类可视化效果。该研究提出的M6模型适用于识别研究区玉米秸秆覆盖区,该方法能够在GEE云平台环境下快速执行,适合推广应用于东北地区秸秆覆盖区。

Crop residue covering is an important way to reduce soil erosion and increase soil organic carbon,which is very important for black soil protection.Therefore,the accurate and rapid identification of corn residue cover area plays an important role in local government monitoring and promoting conservation tillage.The study area is located in Siping City,Jilin Province.Moreover,the time-series Sentinel-2A images collected from GEE(Google Earth Engine)cloud platform are used to capture spectral index based on the characteristics of the corn growing season and after harvest.Index features include Normalized Difference Vegetation Index(NDVI)and Normalized Difference Residue Index(NDRI).The time series feature values are sorted by size,and the quantile method is used to select quartile(QT)features at 0%,25%,50%,75%,and 100%to construct datasets.On this basis,the random forest method after parameter optimization is applied to train and verify the sample datasets divided according to 7∶3,and then the datasets are classified,combined with the connected domain calibration method to remove the small connected domains generated in the classification process,and further optimize the global result.Through the quantitative and qualitative evaluation of Kappa and Overall Accuracy(OA),the experimental results show that:(1)The quantitative evaluation results of the classification model(M1/M2/M3/M4/M5)based on the dataset composed of the different feature are superior 90%.Among them,the classification model M5 of the dataset designed in this paper has the best performance,of which Kappa and OA are 97.41%and 97.91%,respectively.Compared with the classification model M2 without the QT feature,the Kappa and OA are increased by 4.52%and 3.64%,respectively.At the same time,the M5 recognition result can effectively retain edge detail information;(2)For QT feature of different time scales,using the QT feature classification model M5_6/M5 of time series remote sensing images from May to November can greatly restrain another crop residue.Compared with the Kappa and OA of the M5_1 model classification result using only the QT features of the time series images in November,the Kappa and OA increased by 3.9%and 3.12%,respectively;(3)Based on the M5 model,the Kappa and OA of the classification model M6 combined with the connected domain calibration method are 96.76%and 97.36%,respectively,second only to the recognition results of the M5 model.The model M6 avoids fine-grained patches while ensuring high accuracy and optimizing the classification visualization effect.Therefore,the M6 model proposed in this paper is suitable for identifying areas covered by corn residue in the study area.This method can be quickly implemented in the GEE cloud platform environment and is suitable for popularization and application in a corn residue covered areas in Northeast China.