Abstract Accurate spatial information of agricultural parcels is fundamental to any system used in monitoring greenhouse gas emissions, biodiversity developments, and nutrient loading in agriculture. The inefficiency of the traditional methods used in obtaining this information is increasingly paving the way for Remote Sensing (RS). The Multiresolution Segmentation (MRS) algorithm is a well-known method for segmenting objects from images. The quality of segmentation depends on the a priori knowledge of which scale, shape and compactness values to use. With each parameter taking a varied range of input values, this research developed an automated approach for identifying the optimal parameter set without testing all possible combinations. At the core of our approach is Bayesian optimization, which is a sequential model-based optimization (SMBO) method for maximizing or minimizing an objective function. We maximized the Jaccard index, which is a measure that indicates the similarity between segmented agricultural objects and their corresponding reference parcels. As the optimal parameter combination varies between different agricultural landscapes, they were determined at a grid resolution of 10 km. Mono-temporal Sentinel-2 images covering Lower Saxony in Germany were tiled to these grids and the optimal parameters were subsequently identified for each tiled grid. The optimal parameter combinations identified over the grids varied considerably, which indicated that a single parameter combination would have failed to achieve optimal segmentation. We found that the quality of segmentation correlated with the size of agricultural parcels. Under-segmentation was largely minimized but in areas with a predominant agricultural land-use, it was unavoidable. In agricultural parcels composed of heterogeneous pixels, over-segmentation was prevalent. Our approach outperformed other segmentation optimization methods existing in the literature.

中文翻译：

---

基于有监督贝叶斯优化的卫星图像划分农业地块的最优参数

摘要 农业地块的准确空间信息是任何用于监测农业温室气体排放、生物多样性发展和养分负荷的系统的基础。用于获取此信息的传统方法效率低下，正日益为遥感 (RS) 铺平道路。多分辨率分割 (MRS) 算法是一种众所周知的用于从图像中分割对象的方法。分割的质量取决于使用哪个尺度、形状和紧凑度值的先验知识。由于每个参数采用不同范围的输入值，本研究开发了一种自动方法来识别最佳参数集，而无需测试所有可能的组合。我们方法的核心是贝叶斯优化，这是一种基于序列模型的优化 (SMBO) 方法，用于最大化或最小化目标函数。我们最大化了 Jaccard 指数，这是一种指示分割农业对象与其相应参考地块之间相似性的度量。由于不同农业景观之间的最佳参数组合不同，因此它们是在 10 公里的网格分辨率下确定的。覆盖德国下萨克森州的 Mono-temporal Sentinel-2 图像被平铺到这些网格上，随后为每个平铺网格确定了最佳参数。在网格上识别的最佳参数组合变化很大，这表明单个参数组合将无法实现最佳分割。我们发现分割的质量与农业地块的大小相关。欠分割在很大程度上被最小化，但在农业用地占主导地位的地区，这是不可避免的。在由异构像素组成的农业地块中，过度分割很普遍。我们的方法优于文献中现有的其他分割优化方法。