Abstract An algorithm was proposed to classify wheat fields into high-productive and low-productive classes using satellite images according to a threshold value for rain-fed wheat ET a in the anthesis stage. Since the classification process was based on the ET a values in the rain-fed wheat pixels, an algorithm was proposed to map wheat fields using the combination of MODIS and Landsat-8 images. The wheat area mapping method included two major processes in which the first one used a step by step elimination process of non-wheat pixels which did not follow the standard wheat vegetation index time series; the second one used supervised classification methods for detecting the rain-fed wheat pixels. The assessment of the wheat area map was performed statistically using surveyed wheat plots. After detecting the wheat parcels, the threshold value was determined using a frequency analysis on actual evapotranspiration values in the parcels. The rain-fed wheat ET a values estimated from the SEBAL algorithm and were compared to the results from the Eagleman-Affholder method and MOD16A2 products. To assess if the wheat fields in productivity classes have been categorized correctly, yield values in the two classes were compared with each other. The rain-fed wheat yield was estimated using the light use efficiency model and compared to provincial census data for accuracy assessment. Results showed that the overall accuracy, Kappa coefficient, and F1 score values for the 250-m resolution map from the MODIS images were 82, 0.61, and 0.71, respectively. In the same order, these statistics for the 30-m resolution map from the Landsat-8 images were 92, 0.62, and 0.77, respectively. Both the SEBAL and the Eagleman-Affholder methods closely estimated the average wheat ET a value in the anthesis stage equal to 2.4 mm/day. The mean of the rain-fed wheat yield value from the LUE model in 2015 was 10% lower than the census data. Relying on the low amount of the absolute error between the LUE model and the census data, the mean of the yield values in both high-productive and low-productive classes were compared with together. Results showed that the mean amount of yield in high-productive class was 33% (271 kg/ha) more than that of in low-productive class.

中文翻译：

---

利用遥感识别收获前雨养麦田的不同产量潜力

摘要 提出了一种根据雨育小麦开花期ET a 阈值，利用卫星图像将麦田分为高产和低产两类的算法。由于分类过程基于雨养小麦像素中的 ET a 值，因此提出了一种使用 MODIS 和 Landsat-8 图像组合绘制麦田地图的算法。小麦面积制图方法包括两个主要过程，第一个过程使用不遵循标准小麦植被指数时间序列的非小麦像素的逐步消除过程；第二个使用监督分类方法来检测雨养小麦像素。小麦面积图的评估是使用调查的小麦地块进行统计的。检测到小麦包裹后，阈值是使用对地块中实际蒸散值的频率分析确定的。雨养小麦 ET a 值由 SEBAL 算法估计，并与 Eagleman-Affholder 方法和 MOD16A2 产品的结果进行比较。为了评估生产力等级中的麦田是否被正确分类，将两个等级的产量值相互比较。使用光利用效率模型估算雨育小麦产量，并与省级普查数据进行比较以进行准确性评估。结果表明，来自 MODIS 图像的 250 米分辨率地图的整体精度、Kappa 系数和 F1 分数值分别为 82、0.61 和 0.71。按照相同的顺序，来自 Landsat-8 图像的 30 米分辨率地图的这些统计数据分别为 92、0.62 和 0.77。SEBAL 和 Eagleman-Affholder 方法都近似地估计了开花阶段的平均小麦 ET 值等于 2.4 毫米/天。2015 年 LUE 模型的雨养小麦产量平均值比普查数据低 10%。依靠LUE模型与普查数据之间的绝对误差较小，将高产和低产类别的产量平均值进行了比较。结果表明，高生产力等级的平均产量比低生产力等级的平均产量高 33%（271 kg/ha）。依靠LUE模型与普查数据之间的绝对误差较小，将高产和低产类别的产量平均值进行了比较。结果表明，高生产力等级的平均产量比低生产力等级的平均产量高 33%（271 kg/ha）。依靠LUE模型与普查数据之间的绝对误差较小，将高产和低产类别的产量平均值进行了比较。结果表明，高生产力等级的平均产量比低生产力等级的平均产量高 33%（271 kg/ha）。