Abstract Spatially and temporally accurate information on crop areas is a prerequisite for monitoring the multiannual dynamics of crop production. Satellite images have proven their high potential for mapping crop areas at large scales, even at the crop-species level, when a classifier is calibrated on the same image with reference data corresponding to the same period. For operational monitoring purposes, however, it is critical to develop generalized classification methodologies applicable to large scales and different years. Generalized classifiers were presented in this study as follows: a) simple cross-year calibration and application (M1); b) multiyear calibrations (M2); and c) map updating through change detection with multiyear calibrations (M3). These three methods were developed in a classical frame of object-based classifications for a time series of Landsat images with the Random Forest machine learning algorithm. Therein, we tested these methods for sugarcane classification in Sao Paulo state, Brazil, as sugarcane is an economically important crop that has developed substantially in the past decades. Eight years of sugarcane reference maps were used to calibrate and validate the classifiers at four different sites. The cross-year application of M1 provided a low average accuracy Dice coefficient (DC) of 0.84, while it was, on average, 0.94 for the classical same-year calibration. When the classifier was trained on a multiyear dataset (M2), the accuracies achieved average values of 0.91 in independent years. The map updating method M3 showed promising results but was not able to reach the accuracy of visual interpretation methods for detecting annual sugarcane land use change. The multiyear classifier M2 was applied to four contrasting sites and provided reliable results for new sites and years for sugarcane classification. Calibration of the machine learning algorithm on a multiyear dataset of standardized and gap-filled satellite images and reference data proved to give an accurate and space-time generalized classifier, reducing the time, cost and resources for mapping sugarcane areas at large scales.

中文翻译：

---

用于监测巴西甘蔗产区的广义时空分类器

摘要 作物面积的时空准确信息是监测作物生产多年动态的先决条件。当分类器在同一图像上使用对应于同一时期的参考数据进行校准时，卫星图像已经证明它们具有在大尺度上绘制作物区域的巨大潜力，即使是在作物物种级别。然而，出于运行监测的目的，开发适用于大尺度和不同年份的通用分类方法至关重要。本研究中提出的广义分类器如下： a) 简单的跨年校准和应用 (M1)；b) 多年校准（M2）；c) 通过多年校准 (M3) 的变化检测来更新地图。这三种方法是在基于对象分类的经典框架中开发的，用于使用随机森林机器学习算法的 Landsat 图像时间序列。其中，我们在巴西圣保罗州测试了这些用于甘蔗分类的方法，因为甘蔗是一种在过去几十年中取得了长足发展的重要经济作物。八年的甘蔗参考图被用于校准和验证四个不同地点的分类器。M1 的跨年应用提供了 0.84 的低平均精度 Dice 系数 (DC)，而经典同年校准的平均精度为 0.94。当分类器在多年数据集 (M2) 上进行训练时，准确率在独立年份中达到了 0.91 的平均值。地图更新方法 M3 显示出有希望的结果，但无法达到检测年度甘蔗土地利用变化的目视解释方法的准确性。多年分类器 M2 应用于四个对比站点，并为甘蔗分类的新站点和年份提供可靠的结果。机器学习算法在标准化和填补空白的卫星图像和参考数据的多年数据集上的校准证明提供了准确的时空广义分类器，减少了大规模绘制甘蔗区域的时间、成本和资源。