Flooded area classification using pooled training samples: an example from the Chobe River Basin, Botswana
Flooded area classification using pooled training samples: an example from the Chobe River Basin, Botswana
复制标题
DOI：
10.1117/1.jrs.12.026033
复制
发表时间：
2018-06
影响因子：
1.7
通讯作者：
Mitchell P. Braget;D. Goodin;Jida Wang;J. M. Hutchinson;K. Alexander
中科院分区：
文献类型：
--
作者：
Mitchell P. Braget;D. Goodin;Jida Wang;J. M. Hutchinson;K. Alexander
抽象的。在旱地系统中，洪水脉动是系统动力学的驱动力，但流量和景观淹没特征具有很大的变异性。遥感可以为评估和预测水流行为和人口脆弱性提供关键的基础信息；然而，训练和分类大量的时间序列图像是劳动密集型的，限制了这些方法在评估洪水脉冲动力学和景观相互作用方面的有效性。在这里，我们提供了一种替代的方法，该方法只依赖于一组集合的训练样本来进行时间序列图像的分类和分析。我们通过绘制2014年至2016年博茨瓦纳乔贝河流域中分辨率成像光谱仪(MODIS)图像的时间序列来测试这种方法。在汛期(2月至7月)收集的MODIS MOD09A1图像被转换为Kauth-Thomas分量，然后采样形成训练池。然后使用这些汇集的训练样本对图像进行分类。结果表明，使用集合训练获得的分类精度与从常规训练获得的分类在统计上没有区别。将该方法应用于另一年的数据(2013年)也产生了相当准确的结果，这表明训练池方法在应用于图像数据而不是用于创建训练池的图像数据时仍然稳健。
Abstract. In dryland systems, the flood pulse is the driving force in system dynamics but is highly variable in flow volume and landscape inundation features. Remote sensing can provide critical information fundamental to evaluating and forecasting flow behavior and population vulnerability; however, training and classifying an extensive time series of images is labor intensive, limiting the usefulness of these approaches in evaluating flood pulse dynamics and landscape interactions. Here, we provide an alternative approach that relies on only one set of “pooled” training samples for time series image classification and analysis. We test this approach by mapping the flood pulse in a time series of moderate resolution imaging spectroradiometer (MODIS) images from the Chobe River Basin of Botswana for the years 2014 to 2016. MODIS MOD09A1 images collected during the flooding season (February to July) were converted to Kauth–Thomas components, then sampled to form a training pool. Images were then classified using these pooled training samples. Results indicated that classification accuracies obtained using pooled training were statistically indistinguishable from classifications obtained from conventional training. Application of the method to another year’s data (2013) also yielded comparably accurate results, suggesting that the training pool method remains robust when applied to image data other than that used to create the training pool.