Fusion of satellite imagery with different spatial resolutions based on statistical techniques

用于热红外测量的Landsat TM传感器的空间分辨率低于其他六种测量方法。Nishii等人(1996)提出了一种通过将剩余的高分辨率测量线性组合来校正红外测量的方法。然而，基于所有测量的线性组合并不总是提供更好的结果。对于选择问题，Nishii，Kusanobu和Nakaoka(1999)采用了Akaike(1973)的信息论方法。在我们的例子中，AIC被简化为Fujikoshi(1985)所得到的典型相关分析中的AIC。针对这一问题，提出了一种平均AIC方法，并对实际的TM图像进行了分析，取得了较好的效果。与卫星数据分析相关的其他结果如下：-解释了被广泛用于评估土地覆盖分类精度的Kappa统计量不适用于精度评估。交替提出了基于Kullback-Leibler散度的判据。(Nishii和Tanaka，1999)-将森林与人口之间的回归模型扩展到带地貌起伏的空间非线性回归。新模型对广岛的实际数据(田中和西水，1997)进行了很好的拟合。通过空间非线性回归表明，地方政府的肺癌死亡率与距离海岸线的距离有显著的相关性。这表明还需要进一步的流行病学研究。(准备中)-为了去噪，用小波函数对时间序列数据进行展开，并使用贝叶斯模型对小波系数进行收缩。结果表明，该方法能成功地对各种数值算例进行去噪处理。(正在准备中)

The Landsat TM sensor for thermal infrared measurement has lower spatial resolution than that of other six measurements. Nishii et al (1996) proposed a procedure to correct the infrared measurement by a linear combination of remaining high-resolution measurements. However, the linear combination based on all measurements does not always give a better result. For the selection problem, Nishii, Kusanobu and Nakaoka (1999) take an information theoretic approach (AIC) by Akaike (1973). AIC in our case is reduced to that in canonical correlation analysis which is obtained by Fujikoshi (1985). An averaged AIC is proposed for the problem and the procedure shows a good performance for actual TM images.Our other results related to the satellite data analysis are as follows :- It is explained that the kappa statistic, which is widely used for assessing the accuracy of land-cover classification, is not suitable for the accuracy assessment. The criterion based on Kullback-Leibler divergence is alternately proposed. (Nishii and Tanaka, 1999)- The regression model between forest and population is extended to the spatial nonlinear regression with relief as a new covariate. The new model shows a good fit for an actual data of Hiroshima.(Tanaka and Nishui, 1997)- It is shown that the lung cancer mortality in local governments is significantly related to the distance from the shoreline through spatial nonlinear regression. This suggests that the further epidemiological study is required. (in preparation)- For the denoising purpose, time series data is expanded by the wavelet functions and Bayes model is used for shrinking the wavelet coefficients. The procedure is shown to denoise successfully various numerical examples. (in preparation)