Speckle filtering impact on land use/land cover classification area using the combination of Sentinel-1A and Sentinel-2B (a case study of Kirkuk city, Iraq)

Abstract

Providing the information about land use/cover is crucial for many purposes, especially that concern with planning, determination, sprawl and primary human requirements, and health. The extracted information additionally supports in controlling the dynamics of land production from a developing population. In this study, the impact of speckle filtering was examined on the classification of LULC (land use/land cover) by combined data of SAR Sentinel-1A (VH, VV) and multispectral Sentinel-2B data in Kirkuk city, Iraq. Various speckle methods such as Boxcar, Frost, Gamma, and Lee filters were applied on the Sentinel-1A dataset with a different window size and then quantitatively evaluated using various filter performance indices. The Gram–Schmidt (GS) fusion process was chosen to combine the multispectral Sentinel data and VH, VV bands of Sentinel-1 data with different window size. The random forest classifier was applied to classify LULCon, the original data and fused data. The acquired outcomes showed that the Frost filter presented the highest performance of filter using different types of evaluation indices on sentinel-1A (VH, VV) data. The result also demonstrates that the Frost filter is the best filter to enhance the classification precision using the combination data between Sentinal-1(VH) and Sentinel-2B, with an overall accuracy of 97.211% using a 9 × 9 window size and total accuracy of 95.9362% compared with the other filters. Boxcar filter showed the highest classification accuracy indicated in window size 11 × 11, with an accuracy of 95.62%, while the lowest accuracy recorded in 3 × 3 window size of 93.63% with a reduction of efficiency 0.4 % compared with the (VH) data. We obtain good results with increasing of efficiency as 3.191%, 0.192%, 5.582%, 5.58% using windows size of 5 × 5, 7 × 7, 9 × 9, and 11 × 11 respectively, and total enhancement noted was 2.829% compared with the combination with (VV) data.