Calibration and validation data for cropland classification over two sites in 2013

The data are for two sites located in Belgium (top left: 51.00°N, 4.50°E and bottom right: 49.60°N, 5.80°E) and South Africa (top left: 26.85°S, 24.55°E and bottom right: 30.74°S, 29.77°E).. They each cover an area of 60 km x 60 km and are dominated by cropland. For both sites, a 20-m multi-sensor time series consisting of SPOT-4 (Take 5) data and Landsat-8 data spanning from February to December 2013 were at hand. For Landsat, only reflectances in the green, red, near infra-red and short wave infrared wavelengths were considered. The SPOT-4 and Landsat-8 images were processed with the Multi-sensor Atmospheric Correction and Cloud Screening processor. The time series were gap filled using linear interpolation and three spectral-temporal features were extracted. These spectral-temporal features correspond to reflectance composites at the minimum and the maximum of normalized difference vegetation index as well as the mean reflectance over the season. Such features were shown to provide high discrimination power between cropland and non-cropland areas.<br>Ground truth observations from the corresponding growing season supplemented the satellite image time series. In Belgium, field polygons were sourced from the Land Parcel Identification System and non-cropland polygons were digitized based on very high resolution imagery. In South Africa, both cropland and non-cropland objects were both digitized based on very high resolution data. These reference data sets were evenly split at the polygon level into two independent sets to be used for calibration and validation, respectively.

本数据集包含两个研究站点，分别位于比利时与南非。比利时站点的覆盖范围为左上角坐标51.00°N，4.50°E至右下角坐标49.60°N，5.80°E；南非站点的覆盖范围为左上角坐标26.85°S，24.55°E至右下角坐标30.74°S，29.77°E。每个站点的覆盖面积均为60 km × 60 km，且以耕地为主要地物类型。
两个站点均拥有2013年2月至12月期间的20米分辨率多传感器时间序列影像，数据包含SPOT-4（Take 5）与Landsat-8影像。针对Landsat-8影像，仅选取绿波段、红波段、近红外波段与短波红外波段的反射率数据进行分析。所有SPOT-4与Landsat-8影像均通过多传感器大气校正与云筛查处理器（Multi-sensor Atmospheric Correction and Cloud Screening processor）完成预处理。
随后通过线性插值对时间序列数据进行间隙填充，并提取三类光谱-时间特征：分别为归一化差分植被指数（Normalized Difference Vegetation Index, NDVI）最小值与最大值对应的反射率合成数据，以及研究季内的平均反射率。已有研究证实，此类特征可有效区分耕地与非耕地区域。
卫星影像时间序列配套了对应生长季的地面实测标注数据。在比利时区域，耕地地块矢量数据源自土地地块识别系统（Land Parcel Identification System），非耕地地块则通过超高分辨率影像数字化获取；在南非区域，耕地与非耕地对象均通过超高分辨率数据完成数字化。上述参考数据集按照地块级别被均匀划分为两个独立子集，分别用于模型校准与验证。