Titre


Home page




 TEXTUREMOD : processing algorithm


1) A flow of images of same pixel size must be first selected. (No restriction about study sites or sensor type.) For each region considered, a visual delineation of forest areas aims to mask non-forest areas that will be not entered in the FOTO analysis. Efficiency could be reached if permanent masks over rivers, roads, swamps areas, cities are stored independently of specific cloud cover masks that have to be constructed for each image. Resulting masked images are stored within directories of the TEXTUREMOD package and using name coding.
2) From windows size value specified through the graphical interface, the program computes a grid which has as many cells or windows as image dimensions in rows and columns allow.
3) The 2-D FFT is applied to all windows pooling their r-spectra sampled values in a spectra table. Accordingly, it has as many rows as the number of rows matching the number of image windows whereas number of columns remains equal with N/2, the number of sampled frequencies (N/2 corresponds to the Nyquist frequency). Additional information including window location within the image was also stored on supplementary columns so that periodograms of any geographical area could be easily extracted and analyzed. Refer to bibliography for further details.
4) The table of spectra is further submitted to a standardized PCA, that is eigenvector analysis of the correlation matrix between spatial frequencies. The three prominent principal components are used as texture indices (T1, T2, T3) to ordinate windows along texture gradients with a maximum of variance explained.
5) FOTO maps are produced from re-construction of windows grids and by directly applying PC scores as intensity values of a given windows (or pixel). A RGB composite image is then obtained from the first three axes.
6) Thanks to forest parameters table (as given by EXCEL sheet), TEXTUREMOD proposes a basic linear regression analysis of relationships between forest parameters and the three FOTO indices. Inversion relationships is finally applied to the associated FOTO map and a given forest parameter map is automatically produced, respecting all geo-registering information.
Figure 2: Processing algorithm followed by TEXTUREMOD. FP means forest parameter. Click on figure to enlarge.
                           

Performance:

It requires about 15 mn for processing an 16-bits image of 50 Mo using 50-pixels windows with sometimes problems of memory when many images are pooled (it strongly depends on the machine capability). Although Malab® functions are useful for interfacing methods and plotting results, the use of C# language would allow full exploitation of machine memory and windows (forms) interfacing as well. Gain would be obtained in speed and robustness.