The long record of Landsat imagery, which is the cornerstone of Earth observation, provides an opportunity to monitor land use and land cover (LULC) change and understand the interactions between the climate and earth system through time. A few change detection algorithms such as Continuous Change Detection and Classification (CCDC) have been developed to utilize all available Landsat images for change detection and characterization at local or global scales. However, the reliable, rapid, and reproducible collection of training samples have become a challenge for time series land cover classification at a large scale. To meet the challenge, we proposed an automatic phenology learning (APL) method with the assumption that the temporal profiles of samples within the same land cover type are the same or similar at a local scale to generate evenly distributed training samples automatically. We designed the method to build land cover patterns for each category based on consensus samples derived from multiple existing scientific datasets including LANDFIRE's (LF) Existing Vegetation Type (EVT), USGS National Land Cover Database (NLCD), National Agricultural Statistics Service (NASS) Cropland Data Layer (CDL), and National Wetlands Inventory (NWI). Then we calculated the Time-Weighted Dynamic Time Warping (twDTW) distance between any undefined samples and land cover patterns in the same geographical region as prior knowledge. Finally, we selected the optimal land cover category for each undefined sample from the land cover products based on the designed criteria iteratively using the twDTW distance as an indicator. The method was applied in the footprint of 10 selected Landsat Analysis Ready Data (ARD) tiles in the eastern and western conterminous United States (CONUS) to produce annual land cover maps from 1985 to 2017. The accuracy assessment and visual comparison revealed that the APL method can generate reliable training samples without any manual interpretation, producing better land cover results especially for the grass/shrub and wetland land cover classes. Applying the APL method, the overall accuracy of the annual land cover maps was improved by 2% over the accuracy of Land Change Monitoring, Assessment, and Projection (LCMAP) Collection 1.0 Science Products in the research regions. Our results also indicate that the APL method provides an approach for best use of different land cover products and meets the requirement of intensive sampling for training data collection.

Landsat 图像的长期记录是地球观测的基石，它提供了一个机会来监测土地利用和土地覆盖 (LULC) 变化并了解气候和地球系统随时间的相互作用。已经开发了一些变化检测算法，例如连续变化检测和分类 (CCDC)，以利用所有可用的 Landsat 图像在局部或全球尺度上进行变化检测和表征。然而，可靠、快速和可重复的训练样本收集已成为大规模时间序列土地覆盖分类的挑战。为了迎接挑战，我们提出了一种自动物候学习 (APL) 方法，假设同一土地覆盖类型内样本的时间分布在局部范围内相同或相似，以自动生成均匀分布的训练样本。我们设计了基于来自多个现有科学数据集的共识样本为每个类别构建土地覆盖模式的方法，这些数据集包括 LANDFIRE (LF) 现有植被类型 (EVT)、美国地质调查局国家土地覆盖数据库 (NLCD)、国家农业统计局 (NASS)农田数据层 (CDL) 和国家湿地清单 (NWI)。然后，我们计算了与先验知识相同的地理区域中任何未定义样本和土地覆盖模式之间的时间加权动态时间扭曲 (twDTW) 距离。最后，我们根据设计标准迭代地使用 twDTW 距离作为指标，从土地覆盖产品中为每个未定义样本选择最佳土地覆盖类别。该方法应用于美国东部和西部大陆 (CONUS) 的 10 个选定的 Landsat 分析就绪数据 (ARD) 切片的足迹，以生成 1985 年至 2017 年的年度土地覆盖图。精度评估和视觉比较显示 APL方法可以在没有任何人工解释的情况下生成可靠的训练样本，产生更好的土地覆盖结果，特别是对于草/灌木和湿地土地覆盖类。应用 APL 方法，年度土地覆盖图的整体精度比土地变化监测、评估和预测 (LCMAP) 收集 1 的精度提高了 2%。0 研究领域的科学产品。我们的结果还表明，APL 方法提供了一种最佳利用不同土地覆盖产品的方法，并满足了对训练数据收集进行密集抽样的要求。