Abstract Timely and accurate knowledge about the geospatial distribution of crops at regional to continental scales is crucial for forecasting crop production and estimating crop water use. The United States (US) is one of the leading food-producing countries, but lacks a nationwide high resolution crop-specific land cover map available publicly during the current growing season. The goal of this study was to map crops across the Continental US (CONUS) before the harvest, and to estimate the earliest date of classification by which crops can be mapped with sufficient accuracy (90% of full-season accuracy). The study employed a scalable cluster-then-label model that was trained on multiple years of MODIS NDVI using ground truth data in the form of US Department of Agriculture (USDA) Cropland Data Layer (CDL) products. The first step in the crop classification was to perform Multivariate Spatio-Temporal Clustering (MSTC) of annual MODIS-derived NDVI trajectories to create phenologically similar regions, or phenoregions. The second step was to assign crop labels to phenoregions based on spatial concordance between phenoregions and crop classes from CDL using Mapcurves. Assigning crop labels to phenoregions was performed within ecoregions to reduce classification errors due to spatial variability in phenology caused by variations in climate, agricultural practices, and growing conditions. The crop classifier was trained and validated on the years 2008–2014, then tested independently on 2015–2018. Ecoregion-level crop classification performed better than state-level and CONUS-level classification. Pixel-wise accuracy of classification for eight major crops by area was around 70% across the major corn-, soybeans- and winter wheat-producing areas, whereas regions characterized by high crop diversity had slightly lower accuracy. Classification accuracy for dominant crops like corn, soybeans, winter wheat, fallow/idle cropland and other hay/non alfalfa improved with time as they grew, reaching 90% of year-end accuracy by the end of August over each of the four unseen years in the test period. For corn and soybeans, the earliest dates of classification were found to be much earlier in the central regions of the Corn Belt (parts of Iowa, Illinois and Indiana) than in peripheral areas. The ability to map growing crops may permit near real-time monitoring of the health status and vigor of agricultural crops nationally.

中文翻译：

---

绘制美国生长季节内的作物图

摘要 及时准确地了解区域到大陆尺度作物的地理空间分布对于预测作物产量和估算作物用水量至关重要。美国 (US) 是主要的粮食生产国之一，但缺乏当前生长季节公开的全国性高分辨率作物特定土地覆盖图。这项研究的目标是在收获前绘制美国大陆 (CONUS) 的作物图，并估计最早的分类日期，以便能够以足够的准确度（全季准确度的 90%）绘制作物图。该研究采用了一个可扩展的集群然后标签模型，该模型使用美国农业部 (USDA) 农田数据层 (CDL) 产品形式的地面实况数据对多年的 MODIS NDVI 进行了训练。作物分类的第一步是对年度 MODIS 衍生的 NDVI 轨迹进行多变量时空聚类 (MSTC)，以创建物候相似的区域或物候区。第二步是使用 Mapcurves 根据表现区域和来自 CDL 的作物类别之间的空间一致性将作物标签分配给表现区域。在生态区域内为物候区分配作物标签，以减少由于气候、农业实践和生长条件变化引起的物候空间变异引起的分类错误。作物分类器在 2008-2014 年进行了培训和验证，然后在 2015-2018 年进行了独立测试。生态区级别的作物分类表现优于州级别和 CONUS 级别的分类。在主要的玉米、大豆和冬小麦产区，八种主要作物按面积分类的像素精度约为 70%，而作物多样性高的地区精度略低。玉米、大豆、冬小麦、休耕/闲置农田和其他干草/非苜蓿等主要作物的分类准确度随着时间的推移而提高，到 8 月底，在四年未见的年份中的每一年都达到了年终准确度的 90%在测试期间。对于玉米和大豆，玉米带中部地区（爱荷华州、伊利诺伊州和印第安纳州的部分地区）的最早分类日期比外围地区要早得多。绘制作物生长图的能力可能允许近乎实时地监测全国农作物的健康状况和活力。