Automatic extraction of Pseudo-Invariant Features (PIFs) for digital change detection

The process of identifyingthe differences in the status of an object or phenomenon by observing it at different times is called the change detection. In remote sensing change detection, the quantity of a phenomenon is examined from multi-temporal images, and is usually done with the help of multispectral sensors. In remote sensing change detection studies, the type and manner of performing atmospheric correction is one of the most important questions of researchers. In most cases, due to lack of sufficient information or experience and knowledge, absolute radiometric correction is not possible and researchers need to use relative radiometric correction and image-based methods. One of the best image-basedmethods is the radiometric normalization using pseudo-invariant features (PIFs). However, the proper way to select these homogeneous regions remains an important challenge.Therefore, in this research, a very simple method is proposed based on the definition of pseudo-invariant features that automatically identifies these areas and uses a regression process for automatic normalization.

The proposed method in the research is based on the radiometric normalization using pseudo-invariant features. Therefore, it was necessary to identify these areas at first, however, the aim was the automatic extraction of PIFs. According to the definition of pseudo-invariant features, a few basic conditions are needed to define a PIF, therefore, here we have tried to simplify these conditions in order to fall into an automated process:1- Removing water bodies: The study area has a major part of the Persian Gulf coast and water body, which is affected by the tidal wave and under flood conditions; it is affected by the suspended particles of the rivers. Hence, the first step was to remove the water bodies from the images. To mask water from the images, one of the conditions was the pixel value in the NIR band should be less thanthe pixel value in the blue or green band; and another condition was the pixel value in the NIR band should be less the average minus 1 standard deviation of the entire image.2- Removing the areas with vegetation: Generally, in regions with vegetation, the reflectance of the NIR band is higher than RED, therefore, a simple criterion for masking the vegetation is the use of this condition. However, given that the images used in this research are raw and unprocessed, a statistically average was used in this condition. First, the water mask was applied to the images and then, the average of difference of the NIR band and the RED band in the remaining area was obtained. Finally, those areas were selected as vegetation in the whole image,in whichthe difference between these two bands was higher than the calculated average.3- Flatness criterion: The flatness of the area is the simplest criterion for identifying the pseudo-invariant features (PIFs) and is accessible by a digital elevation model with only a slope threshold however, due to the flatness of the study area, this criterion was ignored in this study4- Identifying areas with little or no change over time: In this study, in order to evaluate the effect of radiometric correction in the remote sensing change detection, image algebra change detectionmethodwas used. In this method, spectral image enhancement is done by the use of commonly used spectral vegetation indices. Among the spectral vegetation indices based on the unsupervised classification function, and the measures of the dispersion about the mean of a distribution such as the coefficient of variation, the NDVI index showed a better performance. Accordingly, the NDVI index, which proved to be effective in similar studies, was used further in the analysis. In this index, the NIR band and RED bands are used. Therefore, to identify the unchanged areas, unchanged regions in the NIR and RED bands used in this spectral index were identified and combined. For this purpose, water and vegetation masks were first applied to the multispectral image. Then, the OLI image was stretched to 8 bit to match the ETM + image. In the next step, the difference between the two NIR bands for these two sensors was obtained and the mean value and the standard deviation were calculated. Finally, in order to have the least error, an area was taken into consideration as unchanged area, in which the following relation was present:. The same analysis was done on the red band (). These two criteria were combined together to obtain the unchanged areas by the AND Boolean logic method.
Each one of this four conditions is easy to manually apply to the data with the least processing experience, but in this study, these conditions were automatically generated by the Spatial Model Editor of ERDAS IMAGINE.Radiometric normalization was performed by identifying the pseudo-invariant features (PIFs). In order to validate the accuracy of the proposed method, absolute radiometric correction using ATCOR, FLAASH and ATMOSC methods, and relative radiometric correction using both empirical line calibration method and dark object subtraction method and automatic radiometric correction using QAC and AAIC methods were applied on the data.The output of all atmospheric correction methods and the proposed method was applied in image algebra change detection in the form of a difference and with a threshold of twice the standard deviation from the mean to be checked by 219 points.

The results of validation along with the qualitative studies derived from the histogram comparison proved the proper functioning of the proposed method (Kappa greater than 0.8), and investigating with the help of cross tables indicated that the performance of the proposed method is very similar to that of the empirical line calibration method (More than 76%).

It should be noted that, some unique features in the present research proposal, including simplicity, automation, negligible systematic error, the possibility of using in a biomarker for degradation warning system, the independence on the type of sensor used, differentiate it from other radiometric correction methods, hence, our suggestion to the researchers interested in the remote sensing change detection of the natural ecosystems is to use the findings of this research.