光谱与测光数据融合算法在变星分类上的应用

天文学上把亮度随时间变化的恒星称为变星。它对于研究星系的距离,恒星的演化以及恒星在不同阶段的性质具有非常重要的意义。目前对变星的识别主要依靠长时间观测其亮度变化,并结合对恒星的光谱进行分析才能最终完成认证。这项工作需要天文学家投入大量时间,难以开展大规模分类。针对上述问题本文提出了一种将测光图像与一维光谱进行数据融合用于对变星进行分类的方法——光谱-测光融合网络(ASPF-Net)。该网络由C1网络和C2网络两部分组成,其中C1是用于提取光谱特征的一维卷积神经网络,C2是用于提取测光数据特征的二维卷积神经网络;最后将两者提取到的特征进行融合,用一个全连接前馈神经网络完成分类。该研究在对食变星、脉冲变星和标准星分类问题上进行了实验。实验数据均来自于斯隆数字巡天项目(SDSS),该项目包含了测光图像和光谱两种数据。对于光谱数据本文选取波长在380.0~680.0 nm范围内的流量值。测光图像由:u、g、r、i和z共5个波段数据组成,对应的中心波长分别为:355.1、468.6、616.6、748.0和893.2 nm。相比于传统的利用其中三个波段合成RGB图像,原始SDSS数据拥有更高的灰度等级。为了方便网络训练,对测光数据和光谱数据均做了标准化处理。分类性能分析方面,使用了精确率,召回率,F1值和平均准确率四个指标进行评估。提出的光谱-测光融合网络(ASPF-Net)在针对食双星、脉冲变星和标准星的分类任务,精确率分别为:91.1%、92.8%和98.2%。实验证明,数据融合之后的分类性能优于单独使用光谱数据或测光数据的分类性能。说明将光谱数据和测光数据结合起来对变星进行分类是一种有效的方法,这为今后的变星的分类提供了一种新的思路和方法。

Spatial and Quantitative Comparison of Satellite-Derived Land Cover Products over China

Because land cover plays an important role in global climate change studies, assessing the agreement among different land cover products is critical. Significant discrepancies have been reported among satellite-derived land cover products, especially at the regional scale. Dif- ferent classification schemes are a key obstacle to the comparison of products and are considered the main fac- tor behind the disagreement among the different products. Using a feature-based overlap metric, we investigated the degree of spatial agreement and quantified the overall and class-specific agreement among the Moderate Resolution Imaging Spectoradiometer （MODIS）, Global Land Cover 2000 （GLC2000）, and the National Land Cover/Use Data- sets （NLCD） products, and the author assessed the prod- ucts by ground reference data at the regional scale over China. The areas with a low degree of agreement mostly occurred in heterogeneous terrain and transition zones, while the areas with a high degree of agreement occurred in major plains and areas with homogeneous vegetation. The overall agreement of the MODIS and GLC2000 products was 50.8% and 52.9%, and the overall accuracy was 50.3% and 41.9%, respectively. Class-specific agree- ment or accuracy varied significantly. The high-agreement classes are water, grassland, cropland, snow and ice, and bare areas, whereas classes with low agreement are shru- bland and wetland in both MODIS and GLC2000. These characteristics of spatial patterns and quantitative agree- ment could be partly explained by the complex landscapes, mixed vegetation, low separability of spectro-temporal- texture signals, and coarse pixels. The differences of class definition among different the classification schemes also affects the agreement. Each product had its advantages and limitations, but neither the overall accuracy nor the class-specific accuracy could meet the requirements of climate modeling.

Using multispectral landsat and sentinel-2 satellite data to investigate vegetation change at Mount St. Helens since the great volcanic eruption in 1980

Long-term analyses of vegetation succession after catastrophic events are of high interest for an improved understanding of succession dynamics. However, in many studies such analyses were restricted to plot-based measurements. Contrarily, spatially continuous observations of succession dynamics over extended areas and timeperiods are sparse. Here, we applied a change vector analysis(CVA) to investigate vegetation succession dynamics at Mount St. Helens after the great volcanic eruption in 1980 using Landsat. We additionally applied a supervised random forest classification using Sentinel-2 data to map the currently prevailing vegetation types. Change vector analysis was performed with the normalized difference vegetation index(NDVI) and the urban index(UI) for three subsequent decades after the eruption as well as for the whole observation time between 1984 and 2016. The influence of topography on the current vegetation distribution was examined by comparing altitude, slope angles and aspect values of vegetation classes derived by the random forest classification. WilcoxRank-Sum test was applied to test for significant differences between topographic properties of the vegetation classes inside and outside of the areas affected by the eruption. For the full time period, a total area of 516 km2 was identified as re-vegetated, whereas the area and magnitude of re-growing vegetation decreased during the three decades and migrated closer to the volcanic crater. Vegetation losses were mainly observed in regions unaffected by the eruption and related mostly to timber harvesting. The vegetation type classification reached a high overall accuracy of approximately 90%. 36 years after the eruption, coniferous and deciduous trees have established at formerly devastated areas dominating with a proportion of 66%, whereas shrubs are more abundant in riparian zones. Sparse vegetation dominates at regions very close to the crater. Elevation was found to have a great influence on the reestablishment and distribution of the vegetation classes within the devastated areas showing in almost all cases significant differences in altitude distribution. Slope was less important for the different classes-only representing significantly higher values for meadows, whereas aspect seems to have no notable influence on the reestablishment of vegetation at Mount St. Helens. We conclude that major vegetation succession dynamics after catastrophic events can be assessed and characterized over large areas from freely available remote sensing data and hence contribute to an improved understanding of succession dynamics.

Urban landscape classification using Chinese advanced high-resolution satellite imagery and an object-oriented multi-variable model

The Chinese ZY-1 02C satellite is one of the most advanced high-resolution earth observation systems designed for terrestrial resource monitoring. Its capability for comprehensive landscape classification, especially in urban areas, has been under constant study. In view of the limited spectral resolution of the ZY-1 02C satellite （three bands）, and the complexity and hetero- geneity across urban environments, we attempt to test its performance of urban landscape classification by combining a multi- variable model with an object-oriented approach. The multiple variables including spectral reflection, texture, spatial autocorre- lation, impervious surface fraction, vegetation, and geometry indexes were first calculated and selected using forward stepwise linear discriminant analysis and applied in the following object-oriented classification process. Comprehensive accuracy as- sessment which adopts traditional error matrices with stratified random samples and polygon area consistency （PAC） indexes was then conducted to examine the real area agreement between a classified polygon and its references. Results indicated an overall classification accuracy of 92.63% and a kappa statistic of 0.9124. Furthermore, the proposed PAC index showed that more than 82% of all polygons were correctly classified. Misclassification occurred mostly between residential area and barren/farmland. The presented method and the Chinese ZY-1 02C satellite imagery are robust and effective for urban landscape classification.