Question #2: Noise Reduction | geog4440assignment2

Noise Reduction

In order to make these bands suitable for further data manipulation, filtering and correction must be taken. This section will present all the procedures and techniques of noise removal applied to each band (i.e. Bands 1, 2 and 3). Images before and after correction will be given here for comparison.

Band 1 (Blue): As mentioned in the previous section, medium level of pepper and salt noise, and Artifact #2 at the southwest corner are identified within band 1. In order to deal with these problems, the Algorithm Librarian tool under the 'Tools' tab of Geomatica is the key source of correction. By looking up in Geomatica Online Help, Maximum Noise Fraction noise removal (MNFNR) is a function to remove noise in a single image band using a procedure based on the Maximum Noise Fraction transformation. It is typically suitable for this case where A2Data.pix contains 3 bands (i.e. Bands 1, 2 and 3) which have significantly more noise than the others (i.e. Band 4 and 5), and it is desirable to transform those problematic bands such that their noise content is similar to the other ideal bands.

Before performing MNFNR to band 1, the uncorrected image (i.e. A2Data.pix) was loaded into Geomatica. Meanwhile, uncorrected band 1 was viewed in grey-scale by inputting band 1 into red, green and blue guns at the same time. Then, the Algorithm Library was opened and MNFNR was selected under 'All Algorithms'. After that, MNFNR Module Control panel was popped up. Cleaner bands 4 and 5 were selected as 'Input: Input Image Channels' which act as role models for correction. Noisy band 1 was selected as 'Input Noisy Channel: Channel Containing Noisy Band'. And the 'Input Params 1' was kept as default. After running the function, the corrected image was viewed in Geomatica automatically.

From the comparison above, it is obvious that the salt and pepper noise was significantly diminished. Again, referring back to the lake in the western area, speckles or hazing above were removed. Yet, one of the drawbacks was that the beach or coastal sediment existed in original image was removed at the same time. It was possible that the function of MNFNR classified these pixels into the same class of unwanted speckles, and thus causing this removal. Meanwhile, Artifact #2 was removed successfully. There was only an extremely fainted line left from the shadow of Artifact #2 at the southwest corner of the map. In short, MNFNR is an effective function to remove salt and pepper noise.

Band 2 (Green): As indicated in the previous section, low level of pepper and salt noise, Artifact #1 at the northern part and Artifact #2 at the southwest corner of the map are identified within band 2. At the beginning, I decided to deal with the problem of Artifact #1 first. By looking up in Geomatica Online Help, Image line replacement (LRP) is a algorithm which replaces any line of image data that is presumed damaged in the input image file. The line will be replaced with the line above it, the line below it, or the mean of the lines above and below it.

Before performing LRP to band 2, the uncorrected image (i.e. A2Data.pix) was loaded into Geomatica. Meanwhile, uncorrected band 2 was viewed in grey-scale by inputting band 2 into red, green and blue guns at the same time. After that, the Algorithm Library was opened and LRP was selected under 'All Algorithms'. After that, 'LRP Module Control panel' popped up. Band 2 was selected as 'Input: Input Layer' for correction. And in the tab of 'Input Params 1', 'Replace Mode' was set as 'Mean' while the 'Line Increment Factor' was set as '1'. In order to specify the inserted values for 'X Offset, Y Offset, X Size, Y Size', 'Cursor Control' function was used to find out the coordinates of top left-hand corner and bottom right-hand corner of the Artifact #1 (i.e. dark narrow line). And it was found out that the coordinates are (Pixel: 10, Line: 38) and (Pixel: 830, Line: 39) respectively. Therefore, '10, 38, 830, 1’ was inserted into 'X Offset, Y Offset, X Size, Y Size'. After running the function, the corrected image was viewed in Geomatica automatically.

From the comparison above, it is obvious that Artifact #1 had been removed perfectly. The pixel values of the line removed were filled with the mean of the pixels values above and below them. However, performing LRP only was not enough at this point. Salt and pepper noise, and Artifact #2 still existed across the map. Further correction was needed.

Similar to the steps in the correction of band 1, the Algorithm Library was opened and MNFNR was selected under 'All Algorithms'. After that, MNFNR Module Control panel popped up. Cleaner bands 4 and 5 were selected as 'Input: Input Image Channels' which act as role models for correction. Noisy band 2 (after performing LRP) was selected as 'Input Noisy Channel: Channel Containing Noisy Band'. And the 'Input Params 1' was kept as default. After running the function, the corrected image was viewed in Geomatica automatically.

From the comparison above, it is obvious that the salt and pepper noise was significantly diminished. However, once again, the coastal sediment was misclassified as speckles or hazing and removed accidentally by the MNFNR function. Artifact #2 was removed successfully. There was only a fainted line left from the shadow of Artifact #2 at the southwest corner of the map.

Band 3 (Red): As mentioned in the previous section, high level of pepper and salt noise across the image, and Artifact #2 at the southwest corner are identified within band 3. In order to filer out the noise, both Median Filter (FME, 3x3) and MNFNR filter were performed in this case. By looking up in Geomatica Online Help, FME performs median filtering on image data. The image data will be smoothed while its sharp edges will be preserved.

The Algorithm Library was opened and FME was selected under 'All Algorithms'. After that, FME Module Control panel popped up. The uncorrected band 3 of A2Data.pix was selected as the 'Input: Unfiltered Layer'. 'Input Params 1' was kept as default of which both 'Filter X Size (Pixels)' and 'Filter Y Size (Pixels)' were 3 (i.e. 3x3 Filter). After running the function, the corrected image was viewed in Geomatica automatically.

From the comparison above, it is obvious that the salt and pepper noise was significantly diminished. For example, the noise over the lake in the western area was filtered away. However, there were some areas where the noise was clustered and persisted even after performing FME function. Meanwhile, the presence of Artifact #2 was dampened and fainted, but still existed. The overall image was smoothed and the edges were preserved. After applying FME function, MNFNR function was performed to deal with the remaining noise and Artifact #2.

The Algorithm Library was opened and MNFNR was selected under 'All Algorithms'. After that, MNFNR Module Control panel was popped up. Cleaner bands 4 and 5 were selected as 'Input: Input Image Channels' which act as role models for correction. Noisy band 3 (after performing FME) was selected as 'Input Noisy Channel: Channel Containing Noisy Band'. And the 'Input Params 1' was kept as default. After running the function, the corrected image was viewed in Geomatica automatically.

From the comparison above, it is obvious that the rest of the salt and pepper noise was significantly diminished. Artifact #2 was removed successfully. There was only an extremely fainted line left from the shadow of Artifact #2 at the southwest corner of the map. Unfortunately, coastal sediment was removed as hazing or speckling again accidentally. Last but not least, there was a greater contrast between urban areas and rural areas. And it was favorable for road extraction later on.

At last, corrected bands 1, 2 and 3, and original bands 4 and 5 were merged together by using 'Data Merge' in 'Tools' tab. Projection and resolution in geo-referencing setup were specified from file, which was the original dataset (i.e. A2Data.pix). After executing, the merged image was viewed in Geomatica automatically.

Please Click Image to Enlarge!