A modified temporal criterion to meta-optimize the extended Kalman filter for land cover classification of remotely sensed time series
A modified temporal criterion to meta-optimize the extended Kalman filter for land cover classification of remotely sensed time series
复制标题
DOI：
10.1016/j.jag.2017.12.007
复制
发表时间：
2018-05
期刊：
影响因子：
--
通讯作者：
B. P. Salmon;W. Kleynhans;J. Olivier;F. V. D. Bergh;K. Wessels
中科院分区：
文献类型：
--
作者：
B. P. Salmon;W. Kleynhans;J. Olivier;F. V. D. Bergh;K. Wessels
人类正在以越来越快的速度改变土地覆盖。准确的土地覆盖地理地图，特别是农村和城市住区，对规划可持续发展至关重要。从中分辨率成像光谱仪(MODIS)地表反射率产品中提取的时间序列通过分析反射率值的季节模式来区分土地覆盖类别。为了使参数模型适合这些时间序列，通常需要对回归方法进行多次调整。为了减少工作量，对地理区域的回归方法进行了参数的全局设置。在这项工作中，我们修改了一种元优化方法，以设置回归方法来提取每个时间序列的参数。以模型参数的标准差和残差的大小作为评分函数。我们使用非线性扩展卡尔曼滤波(EKF)成功地将一个三重调制模式与我们研究区域的季节模式进行了拟合。该方法利用时间信息，大大减少了处理每个时间序列的处理时间和存储需求。它还分别得出每个时间序列的可靠性度量。利用支持向量机对提取的特征进行分类，并与原方法在地面真实数据上的性能进行了比较。
Humans are transforming land cover at an ever-increasing rate. Accurate geographical maps on land cover, especially rural and urban settlements are essential to planning sustainable development. Time series extracted from MODerate resolution Imaging Spectroradiometer (MODIS) land surface reflectance products have been used to differentiate land cover classes by analyzing the seasonal patterns in reflectance values. The proper fitting of a parametric model to these time series usually requires several adjustments to the regression method. To reduce the workload, a global setting of parameters is done to the regression method for a geographical area. In this work we have modified a meta-optimization approach to setting a regression method to extract the parameters on a per time series basis. The standard deviation of the model parameters and magnitude of residuals are used as scoring function. We successfully fitted a triply modulated model to the seasonal patterns of our study area using a non-linear extended Kalman filter (EKF). The approach uses temporal information which significantly reduces the processing time and storage requirements to process each time series. It also derives reliability metrics for each time series individually. The features extracted using the proposed method are classified with a support vector machine and the performance of the method is compared to the original approach on our ground truth data.