Shallow water depth is one of the important factors in science and coastal environmental management. However, in-situ measurement is quite costly and time-consuming. Past research efforts have provided a number of optically-based methods to estimate shallow water depth distribution from satellite image, but they cannot properly handle with the heterogeneity in bottom sediment distribution because they require image-specific assumptions or additional information on bottom reflectivity. It is therefore indispensable to develop a method that can be applied more generally to water areas with inhomogeneous bottom material. In any application of depth prediction methods, we need depth data for some points to validate the results. A leave-one-out cross validation technique enables us to use the data for predictive model building without degrading the reliability of prediction error evaluation. From this standpoint, we present a new generalized method over the previous methodologies by utilizing depth measurement data. In the new method, the bottom reflection term of the optical model is assumed to be a nonparametric function of the depth-independent variables (bottom indexes), which can be calculated from the brightness values of the pixels. In this way, the water depth is explained by a semiparametric regression model. The ratios of the diffuse attenuation coefficients, which are needed to calculate the bottom index, are optimized to minimize Generalized Cross-Validation(GCV). The new method is applied to 3 coral reef areas and artificially generated situations, and the prediction accuracy is compared with those of the methods proposed by Paredes et al., Stumpf et al., and Kanno et al. As a result, the new method is found to have the highest accuracy in cases that enough depth-known pixels are available and that the optical model apply well.
A new topographic correction method was developed, which can be applied to satellite image in the season of low solar elevation. At first, Minnaert method and C method were tested with the ASTER image of the mountainous area for the time when solar elevation was low to clarify problems of these methods. As a result, Minnaert method could not correct the part where the cosine of the solar incidence angle (cos i) was smaller than or equal to zero because of structural problem of the correction formula. In C method, correction effect was insufficient in the part where cos i was positive. Also, overcorrections were found in the part where cos i was negative because distribution characteristics of Digital Number (DN) were different between positive part and negative part of cos i. In addition, in the winter image used in this study, original DN value had elevation dependence caused by colored leaves and fallen leaves. In order to solve these problems, DPR (Dual Partitioning Regression) method was developed. In this method, sample data were extracted from each land cover and elevation. DPR method uses inclination of the linear regression line of cos i versus original brightness Do as the correction parameter. In this regard, regression was calculated in positive part and negative part of cos i separately. The correction formulas were derived in such a way that corrected brightness Dc became equal to original brightness Do when cos i was 1. Topographic correction by DPR method was performed with the image which Minnaert method and C method were applied to. As a result, the coefficient of correlation for regressions between cos i and Dc showed a very low value, less than 0.03, in all bands.
In satellite remote sensing, higher spatial resolution than the particular resolution of a sensor may be achieved by observing a same area repeatedly at an interval decided by the recurrent interval days. The methods have been researched to improve the spatial resolution by fusion of multiple low-resolution images and by the multiple observations of subpixel-shifted low-resolution images. These methods are known as the super-resolution (SR) reconstruction. Various methods of the SR reconstruction have been suggested till now. When resolution is raised N times by the SR reconstruction, peculiar subpixel-shifted images of about N×N numbers or even more are required. Uncontrollable misregistration by slight orbital translation may often be present in the observed images. And, it can be actually obtained limited pieces of observed images under a same weather condition in a same season. In this paper, a new SR reconstruction method is proposed by using a number of incomplete low-resolution images with random positional shifts. This proposed method enables sufficient SR performance by giving the preliminally process such as re-arrangement of elements of low-resolution images on the continuous space and supplementation with insufficient low-resolution images by the 2-dimensional interpolation. The simulation using sample images and the experiment using actual images taken in the laboratory shows effectiveness.
Illegal dumping of industrial waste causes severe environmental damage. For minimizing the damage, it is important to find out an occurrence of illegal dumping in its early stage. However, illegal dumping often occurs in area where public notice doesn't reach. Illegal dumping in such area escapes from the surveillance of local governmental agents, then its scale tends to become larger. Satellite images can provide useful information to the agents for finding such the hidden illegal dumping since it can observe land surface conditions in extensive area from the above. The purpose of this study is to detect illegal dumping sites in an acquired satellite image by image analysis for supporting visual check by the agents. We attempt to develop a practical detection method that can be applied to more various kinds of illegal dumping sites. To realize this detection we conducted hearing investigations to the agents and on-site investigations. As a result of the investigations, it has been understood that information of vegetation removal and landscape structure change are useful for the detection. In this paper we proposed the detection method that uses mainly the information of vegetation removal and land surface structure change. In addition, application of the proposed method to actual images was examined.