Levitan Nathaniel, Gross Barry
Department of Electrical Engineering, City College of New York, 160 Convent Ave., New York, NY 10031, USA;
Remote Sens (Basel). 2018;10(12):1968. doi: 10.3390/rs10121968. Epub 2018 Dec 6.
Due to its worldwide coverage and high revisit time, satellite-based remote sensing provides the ability to monitor in-season crop state variables and yields globally. In this study, we presented a novel approach to training agronomic satellite retrieval algorithms by utilizing collocated crop growth model simulations and solar-reflective satellite measurements. Specifically, we showed that bidirectional long short-term memory networks (BLSTMs) can be trained to predict the in-season state variables and yields of Agricultural Production Systems sIMulator (APSIM) maize crop growth model simulations from collocated Moderate Resolution Imaging Spectroradiometer (MODIS) 500-m satellite measurements over the United States Corn Belt at a regional scale. We evaluated the performance of the BLSTMs through both k-fold cross validation and comparison to regional scale ground-truth yields and phenology. Using k-fold cross validation, we showed that three distinct in-season maize state variables (leaf area index, aboveground biomass, and specific leaf area) can be retrieved with cross-validated R values ranging from 0.4 to 0.8 for significant portions of the season. Several other plant, soil, and phenological in-season state variables were also evaluated in the study for their retrievability via k-fold cross validation. In addition, by comparing to survey-based United State Department of Agriculture (USDA) ground truth data, we showed that the BLSTMs are able to predict actual county-level yields with R values between 0.45 and 0.6 and actual state-level phenological dates (emergence, silking, and maturity) with R values between 0.75 and 0.85. We believe that a potential application of this methodology is to develop satellite products to monitor in-season field-scale crop growth on a global scale by reproducing the methodology with field-scale crop growth model simulations (utilizing farmer-recorded field-scale agromanagement data) and collocated high-resolution satellite data (fused with moderate-resolution satellite data).
由于其全球覆盖范围和高回访时间，基于卫星的遥感技术能够在全球范围内监测作物生长季的状态变量和产量。在本研究中，我们提出了一种新颖的方法，通过利用并置的作物生长模型模拟和太阳反射卫星测量数据来训练农艺卫星反演算法。具体而言，我们表明，可以训练双向长短期记忆网络（BLSTM），根据美国玉米带地区尺度上并置的中分辨率成像光谱仪（MODIS）500米卫星测量数据，预测农业生产系统模拟器（APSIM）玉米作物生长模型模拟的生长季状态变量和产量。我们通过k折交叉验证以及与区域尺度地面真值产量和物候的比较，评估了BLSTM的性能。使用k折交叉验证，我们表明，在生长季的大部分时间里，可以反演三个不同的生长季玉米状态变量（叶面积指数、地上生物量和比叶面积），交叉验证的R值范围为0.4至0.8。本研究还通过k折交叉验证评估了其他几个植物、土壤和物候生长季状态变量的可反演性。此外，通过与基于调查的美国农业部（USDA）地面真值数据进行比较，我们表明，BLSTM能够预测实际县级产量，R值在0.45至0.6之间，以及实际州级物候日期（出苗、抽丝和成熟），R值在0.75至0.85之间。我们认为，该方法的一个潜在应用是通过使用田间尺度作物生长模型模拟（利用农民记录的田间尺度农业管理数据）和并置的高分辨率卫星数据（与中分辨率卫星数据融合）来重现该方法，从而开发卫星产品，在全球范围内监测田间尺度的作物生长季生长情况。
