Abstract Multi-sensor remote sensing data fusion technologies have been developed and widely applied in recent years, providing a feasible and economical solution to increase the availability of high spatial and temporal resolution data. These methods, however, have been challenging to apply in highly heterogeneous areas, especially in complex agricultural landscapes where there are rapid changes at small scales, while features at larger scales change more slowly. In this study, we developed a novel method to reconstruct daily 30 m Normalized Difference Vegetation Index (NDVI) using imagery from the Moderate Resolution Imaging Spectroradiometer (MODIS), Landsat and Landsat-like platforms, and the Cropland Data Layer (CDL). This method utilizes a crop reference curve (CRC) approach, in which a set of NDVI time series are extracted from pure MODIS pixels (250 m resolution) identified using the CDL, and then used to fit Landsat-like observations (30 m). The CRC based method was applied over a complex agricultural landscape in the Choptank River watershed on the Eastern Shore of Maryland. Landsat data from 2013 and 2014 and Harmonized Landsat and Sentinel-2 (HLS) data from 2018 were used to reconstruct 30 m daily NDVI maps for major crop types. Results show that the relative error (RE) in reconstructed NDVI is around 6–8% during periods of rapid crop growth, and 3–5% during peak periods when growth is slow. The accuracy of the CRC method outperforms a standard image pair-based data fusion algorithm (Spatial and Temporal Adaptive Reflectance Fusion Model; STARFM), which yields RE of 4–9% in slow-growth periods and 10–16% in fast-growth periods when clear Landsat images are scarce. The CRC method was also compared with time-series data fusion methods, including a harmonic fitting model and the SaTellite dAta IntegRation (STAIR) model. The results show that CRC gives similar results when the Landsat-like image availability is high (around 27 images per year), but outperforms other methods when availability is limited (less than 15 images per year). The reconstructed NDVI time series for corn, soybean, winter wheat/soybean and forest at 30-m resolution show clear phenological patterns at the sub-field scale. The resulting 30-m NDVI timeseries data provide useful information for mapping crop phenology and monitoring crop condition in complex agricultural landscapes, especially for complex double-cropping areas. However, the input requirement of an accurate 30-m crop classification map constrains its application to areas and periods where classifications are available.

中文翻译：

---

使用作物参考曲线方法重建复杂农业景观的每日 30 m NDVI

摘要 近年来，多传感器遥感数据融合技术得到发展和广泛应用，为提高高时空分辨率数据的可用性提供了一种可行且经济的解决方案。然而，这些方法在高度异质的地区应用具有挑战性，特别是在小尺度变化迅速而大尺度特征变化较慢的复杂农业景观中。在这项研究中，我们开发了一种使用中分辨率成像光谱仪 (MODIS)、Landsat 和类陆地卫星平台以及农田数据层 (CDL) 的图像重建每日 30 m 归一化差异植被指数 (NDVI) 的新方法。该方法利用作物参考曲线 (CRC) 方法，其中从使用 CDL 识别的纯 MODIS 像素（250 m 分辨率）中提取一组 NDVI 时间序列，然后用于拟合类 Landsat 观测（30 m）。基于 CRC 的方法应用于马里兰州东岸乔普坦克河流域的复杂农业景观。使用 2013 年和 2014 年的 Landsat 数据以及 2018 年的协调 Landsat 和 Sentinel-2 (HLS) 数据重建主要作物类型的 30 m 每日 NDVI 地图。结果表明，重建的 NDVI 的相对误差 (RE) 在作物快速生长期间约为 6-8%，在生长缓慢的高峰期约为 3-5%。CRC 方法的准确性优于基于标准图像对的数据融合算法（Spatial and Temporal Adaptive Reflectance Fusion Model；STARFM），当清晰的 Landsat 图像稀缺时，在缓慢增长时期产生 4-9% 的可再生能源，在快速增长时期产生 10-16% 的可再生能源。CRC 方法还与时间序列数据融合方法进行了比较，包括谐波拟合模型和卫星数据集成 (STAIR) 模型。结果表明，当类似 Landsat 的图像可用性很高（每年大约 27 张图像）时，CRC 给出了类似的结果，但在可用性有限（每年少于 15 张图像）时优于其他方法。以 30 米分辨率重建的玉米、大豆、冬小麦/大豆和森林的 NDVI 时间序列在子场尺度上显示出清晰的物候模式。由此产生的 30 米 NDVI 时间序列数据为绘制作物物候和监测复杂农业景观中的作物状况提供了有用的信息，特别是对于复杂的双季作物地区。