Mapping and monitoring of urban land cover are becoming increasingly important in the face of ongoing climate change. The recently launched pair of Sentinel-2 satellites regularly delivers images with the highest spectral and spatial resolution among the freely available images. Based on these data, we proposed an object-based classification method for urban land-cover mapping at two basic scale levels. The proposed set of classification rules is transferable to different urban areas without the need to collect training samples. The inevitable problem of different spectral characteristics of vegetation in individual areas is solved by computing the area-specific thresholds using the central values of forest stands and grasslands read out of the histogram. Special features summarising the normalised difference vegetation index (NDVI) time-series using the Google Earth Engine platform were designed to distinguish cropland from other classes. The transferability of the rule sets was verified in six Central European cities with different climatic conditions—Bratislava, Nitra, and Žilina (Slovakia), Zakopane (Poland), and Kaposvár and Orosháza (Hungary). The overall classification accuracy reached 78–90% in each tested area at the first hierarchical level, and 76–89% on the second level, respectively. The performance of the methodology was compared with the random forest (RF) machine learning method with training samples collected in Bratislava. The results confirmed that without area-specific training samples, the accuracy of the RF method is 5–35 percentage point (p.p.) lower than the accuracy of the proposed rule-based method. In addition, without the NDVI summary indices, the accuracy of the RF classifier decreased by another 10–30 p.p.

面对持续的气候变化，对城市土地覆盖进行制图和监测变得越来越重要。最近发射的一对Sentinel-2卫星定期提供免费提供的图像中具有最高光谱和空间分辨率的图像。基于这些数据，我们在两个基本尺度上提出了一种基于对象的城市土地覆盖制图分类方法。提议的分类规则集可以转移到不同的城市区域，而无需收集培训样本。通过使用从直方图中读出的林分和草地的中心值来计算特定于区域的阈值，可以解决各个区域中植被光谱特征不同的不可避免的问题。使用Google Earth Engine平台汇总归一化差异植被指数（NDVI）时间序列的特殊功能旨在将农田与其他类别区分开来。在六个具有不同气候条件的中欧城市中验证了规则集的可传递性，它们分别是布拉迪斯拉发，尼特拉和Žilina（斯洛伐克），扎科帕内（波兰）以及卡波什瓦尔和奥罗萨扎（匈牙利）。在第一层次上，每个测试区域的总体分类准确度分别达到78-90％，在第二层次上，分别达到76-89％。该方法的性能与在布拉迪斯拉发收集的训练样本的随机森林（RF）机器学习方法进行了比较。结果证实，在没有特定区域训练样本的情况下，RF方法的准确性为5–35个百分点（pp ）低于提出的基于规则的方法的准确性。此外，如果没有NDVI摘要索引，RF分类器的精度又降低了10–30 pp