基于非线性降维时序遥感影像的作物分类

当前基于时序遥感数据的作物分类方法大都需要较多专家知识及人工干预,难以自动化,也难以移植到其他地区。将光谱降维技术用于时序遥感影像分析可以很好地解决这一问题。其中,非线性降维方法已经成功应用于高光谱数据,并且获得了比线性降维方法更好的结果。但是,直接将非线性降维方法用于时序遥感影像无法充分利用其时相维度的信息。该文改进了一种非线性降维算法——Laplacian Eigenmaps(LE)用于时序遥感影像的作物分类,该方法更加关注相同时相下不同作物生长季的物候特征差异,而不再仅依赖于整个生长季的物候曲线轮廓。改进的LE算法被应用于美国伊利诺伊州覆盖作物全生长季的Landsat 8时间序列影像。降维后保留的波段结合随机森林分类器基于美国农业部Cropland Data Layer(CDL)提供的训练数据完成了一系列的分类试验,并与传统插值未降维的方法进行对比。试验结果表明,改进的LE降维方法完成了更高的整体及各个类别的分类精度,其中整体分类精度达到85.37%,该方法作为一种自动化的方法,不需要人工干预,可直接移植到其他研究区,并且只需要较少的训练样本就可以完成一个较高的分类精度,为日后不同尺度的作物识别和提取研究提供了有效的方法。

Crop classification based on nonlinear dimensionality reduction using time series remote sensing images

Abstract: Compared with the traditional method of ground statistics, remote sensing technology has the unique advantages of high efficiency, fast speed and high precision. It is suitable for agricultural monitoring, especially the time series remote sensing data can provide spectral information covering the whole growing season of crops. At present, most of the crop classification methods based on time series remote sensing data need many expert knowledge and manual intervention, which are difficult to automate and transplant to other areas. The application of dimensionality reduction (DR) technique to time series remote sensing image analysis can solve this problem efficiently. The imaging process of remote sensing image will experience multiple scattering between different scene components. This scattering is influenced by many factors such as wavelength, observation angle, illumination condition, and three-dimensional structure of scene component. Therefore, the remote sensing image has inherent nonlinear characteristics, and the nonlinear DR technology is more suitable for remote sensing images. The nonlinear spectral DR method has been successfully applied to hyperspectral data and obtained better results than linear DR methods. However, the nonlinear DR method cannot make full use of the temporal information in time series remote sensing images. In this study, a nonlinear DR algorithm, Laplacian Eigenmaps (LE), is refined for crop classification in time series remote sensing images. This method emphasizes the phenological characteristics difference of different crop growing seasons. In the refined LE algorithm, the dynamic time warping (DTW) measurement with different weight value is used to calculate the similarity between the time series. If the two phases of the calculated path are consistent or close, a smaller weight should be given, and a greater weight should be given when the two phases correspond to the path are far away. In addition, the refined LE method can be applied to time series data with unequal length and reduce the preprocessing operation of time series data. Moreover, the proposed method is simple to run and only requires two input parameters. One is the midpoint of time series, and the other is the number of retained bands after DR. Both two parameters are easily to obtain. The refined LE algorithm is applied to the Landsat time series images of the whole growing season of crops in Illinois, USA. A series of classification experiments based on the training data provided by the United States Department of Agriculture National Agricultural Statistics Services Cropland Data Layer (CDL) are completed and compared with the traditional interpolation method without DR and the LE-DTW method proposed for land cover classification. The experimental results show that the refined LE DR method has completed a higher classification accuracy and per-class accuracy than other methods. And the overall classification accuracy of the proposed method reaches 85.37%, which is 9.05 and 3.45 percentage points higher than that of temporal interpolation (TI) method and LE-DTW method respectively. And the method can achieve a given degree of classification accuracy with only a small number of training samples. When the proportion of training samples is higher than 1%, the positive and negative standard deviations of the proposed method is below 1%. The overall classification accuracy of the proposed method tends to be stable when the proportion of training samples reaches 2%. The refined LE dimensionality reduction method is an automatic method and can be directly applied to other research areas without manual intervention. It provides an effective method for crop identification and extraction on different scales in the future.