基于不同时间尺度与特征优选的黄淮海平原冬小麦识别

Identification of Winter Wheat in Huang-Huai-Hai Plain Based on Different Time Scales and Feature Preference

摘要:

准确及时地监测区域作物种植面积对保障我国粮食安全和农业可持续发展具有重要意义。本研究利用Google Earth Engine（GEE）云平台和融合的Sentinel-1 SAR雷达影像与Sentinel-2 SR地表反射率影像，对黄淮海平原2021年冬小麦进行了监督分类。通过对Sentinel时间序列数据进行不同时间尺度合成与平滑处理，并对极化特征、光谱特征、植被指数、谐波系数和纹理特征进行优选，以探究不同时间尺度的影像序列及特征优选对黄淮海平原冬小麦识别精度和泛化能力的影响。结果表明：特征优选过程可以提高模型分类精度，在各类特征因子中，光谱特征重要性最高，其次为谐波系数、极化特征和纹理特征。随着影像序列时间尺度的缩减，可以得到更高的分类精度，尺度30、20、10d平均总体精度分别为95.4%、95.6%和96.4%；但泛化能力也随之降低，对应的泛化能力分别为0.935、0.919和0.918。短时间尺度影像序列能够更准确地捕获地物的特征细节，展现出更高的分类精度，但其对数据变化的适应能力更差。此外，模型泛化能力在空间上呈现“越近越相关”的规律。利用GEE平台及Sentinel系列卫星遥感数据，实现了对黄淮海平原冬小麦面积的准确识别。整体上，混淆矩阵总体精度（OA）和F1值均在90%以上，分类结果在空间细节上与高分辨率图像高度一致，同时提取的冬小麦种植面积与市级官方统计数据高度相关（决定系数R2>0.9）。

关键词:冬小麦;遥感识别;Google Earth Engine;随机森林;时间尺度;特征优选

Abstract:

Monitoring regional crop acreage accurately and promptly is critical for ensuring food security and promoting sustainable agricultural development in China. The Google Earth Engine (GEE) cloud platform, along with fused Sentinel-1 SAR radar and Sentinel-2 SR surface reflectance imagery were employed to classify winter wheat in 2021 within the Huang-Huai-Hai Plain. The Sentinel time series data were synthesized and smoothed across various temporal scales, and a prioritization of polarization features, spectral features, vegetation index, harmonic coefficients, and textural features were conducted to explore their impacts on the accuracy and generalization ability of winter wheat identification in the region. The results showed that the feature optimization process improved the classification accuracy of the model, and the spectral features were the most significant, followed by harmonic coefficients, polarization, and textural features. Reducing the time scale of image sequences led to higher classification accuracy, with overall accuracies of 95.4%, 95.6%, and 96.4% for 30 d, 20 d and 10 d scales, respectively. However, this also resulted in a decrease in generalization ability, with corresponding scores of 0.935, 0.919, and 0.918. Shorter time scales captured finer details of ground features, achieving higher classification accuracy but showing less adaptability to data variations. Moreover, the model’s generalization ability demonstrated a spatial pattern of ‘the closer it was, the more relevant they were’. The identification of winter wheat areas using the GEE platform and Sentinel imagery was highly accurate, with overall accuracy and F1 scores of the confusion matrix exceeding 90%, and classification results were highly consistent in spatial detail with high-resolution images. Furthermore, the estimated areas of winter wheat showed a strong correlation with official municipal statistics (coefficient of determination R2>0.9).

Key words:winter wheat;remote sensing imagery identification;Google Earth Engine;random forest;time scale;feature preference