Performance evaluation of principal component analysis, independent component analysis and minimum noise fraction method in increasing the information extracting accuracy of Sentinel-2 satellite data

Subject Areas : Geospatial systems development

Sayyad Asghari Saraskanrood ¹ , Hasan Hasani Moghaddam ² , Hossein Fekrat ³

1 - Associate Professor, Department of Natural Geography, Faculty of Humanities, University of Mohaghegh Ardabili, Iran
2 - MSc. of Remote Sensing and GIS, Faculty of Humanities, Kharazmi University, Tehran, Iran
3 - MSc. Student of Remote Sensing and GlS, Faculty of Humanities, University of Mohaghegh Ardabili, Iran

Received: 2020-09-02 Accepted : 2020-11-02 Published : 2021-06-22

Keywords: transformation, Rezvanshahr, Sentinel-2, Maximum likelihood, Sheffield,

Abstract :

Background and ObjectiveProblem The use of various transformations to improve the accuracy of data extraction from satellite images is increasing sharply. In the meantime, the choice of optimal conversion is very important and will affect the output results. Due to the correlated nature of remote sensing images, the use of various transformations to improve the accuracy of information extraction from these images is essential. According to the studies, the purpose of this study is to investigate different methods of image conversion in improving the process of classification of satellite images and increasing the accuracy of land use maps. Considering that the study area and in general the northern regions of Iran are facing special conditions of entanglement of land uses, so the use of various conversion methods as well as the combined method proposed in this study increases the accuracy and the accuracy of the output information and finally the possibility of more detailed separation and review of uses and identification of factors changing them for future planning.Materials and Methods In this study, in order to evaluate the performance of principal component analysis methods, independent component analysis, and minimum noise fraction method, Sentinel-2 satellite images of Rezvanshahr city were used. Gram-Schmit algorithm was used to integrate this data with each other and achieve a resolution of 10 meters. After applying the necessary pre-processing and merging the images together, all three transformations were applied to the image, as well as a combination of the components of these three methods. Then, the results of the transformations were classified into 8 user classes using the maximum likelihood algorithm. Using Sheffield coefficient and statistical calculations between the obtained components, the combination of the first components of principal component analysis, the first component of minimum noise fraction, and the second component of independent component analysis were selected as the optimal combination. General knowledge of the area and accordingly the visual interpretation of the outputs, as well as the perception of 120 ground points by GPS, has been the basis for assessing the accuracy of the output maps.Results and Discussion After applying the required preprocessors, each of these algorithms was applied to the image, and the output of each was classified into 8 user classes using the Maximum Likelihood algorithm. The results of output maps showed that the conversion of principal component analysis, considering that it considers Gaussian distribution for variables and tries to decompose the extracted components, is weak in samples with non-Gaussian distribution and shows low performance. The minimum noise fraction algorithm works similarly to the principal component analysis algorithm, except that it classifies the noise better. This algorithm has less error in separating classes and this factor has resulted in better performance and higher accuracy than the other two conversions. In the independent component analysis algorithm, the image correlated bands of the study area have been converted to independent components and new information has been extracted from the area. The visual interpretation shows the high accuracy of the classification result and an error matrix (confusion) is used to quantify the accuracy of the classified image. The results of the evaluation of overall accuracy and kappa coefficient showed that the classification of the original image without applying transformations and with the same training samples of output with an overall accuracy of 76% and kappa coefficient of 0.78 had the highest error. Also, the results of other outputs for classification resulting from principal component analysis conversion are 80% overall accuracy and kappa coefficient of 0.83, respectively, for classification resulting from minimum noise fraction conversion, total accuracy of 85% and kappa coefficient of 0.88 and for the classification obtained from the analysis of independent component analysis, the overall accuracy was 77% and the kappa coefficient was 0.80. After selecting the optimal combination of components of principal components analysis methods, independent component analysis and minimum noise fraction method and selecting the first components of principal component analysis algorithms and minimum noise fraction and the second component of total component analysis to 92% independent coefficient and Kappa increased 0.94.Conclusion In this study, after evaluating the conversion performance of principal component analysis, independent component analysis, and minimum noise fraction method, an optimal combination of components of these transformations was proposed. As the results of the research showed, the classification of the original image without conversions and with the same training samples had low overall accuracy and kappa coefficient. The results show the close performance of these transformations to each other, which indicates the existence of both Gaussian and non-Gaussian distributions of variables. MNF conversion has minimized the amount of data noise and results in better output than ICA and PCA conversion. Since these transformations alone are not able to extract all the components of the image, so a combination of the components of these transformations based on the Sheffield coefficient was chosen to assume the Gaussian and non-Gaussian distributions of the variables with the least possible noise.

References:

Al-Ahmadi F, Al-Hames A. 2009. Comparison of four classification methods to extract land use and land cover from raw satellite images for some remote arid areas, Kingdom of Saudi Arabia. Earth Sciences, 20(1): 167-119.

Alikhah-Asl M, Elham F, Mohammad N. 2014. Evaluation of different enhancement remote sensing techniques. International Journal of Agriculture Innovations and Research, 3(1): 33-37.

Anderson JR. 1971. Land-use classification schemes. Photogrammetric Engineering, 37(4): 379-387. https://trid.trb.org/view/93641.

Arslan O, Akyürek Ö, Kaya Ş. 2017. A comparative analysis of classification methods for hyperspectral images generated with conventional dimension reduction methods. Turkish Journal of Electrical Engineering & Computer Sciences, 25(1): 58-72. doi:https://doi.org/10.3906/elk-1503-167.

Barzegari Dehaj V, Zare M, Mokhtari MH, Ekhtesasi MR. 2018. Evaluation of different satellite image enhancement techniques in separating of geological units. Journal of RS and GIS for Natural Resources, 9(1): 1-23. http://girs.iaubushehr.ac.ir/article_540405_en.html. (In Persian).

Bellvert J, Jofre-Ĉekalović C, Pelechá A, Mata M, Nieto H. 2020. Feasibility of using the two-source energy balance model (TSEB) with Sentinel-2 and Sentinel-3 images to analyze the spatio-temporal variability of vine water status in a vineyard. Remote Sensing, 12(14): 2299. doi:https://doi.org/10.3390/rs12142299.

Congalton RG, Green K. 2019. Assessing the accuracy of remotely sensed data: principles and practices. CRC press. 139 p.

Dabiri Z, Lang S. 2018. Comparison of independent component analysis, principal component analysis, and minimum noise fraction transformation for tree species classification using APEX hyperspectral imagery. ISPRS International Journal of Geo-Information, 7(12): 488. doi:https://doi.org/10.3390/ijgi7120488.

ESA. 2017. (Standard Document), SENTINEL-2 User Handbook, 2.

ESA. 2018. SNAP-Sen2Cor, Available online: http://step.esa.int/main/third-party-plugins-2/sen2cor.

Estornell J, Martí-Gavilá JM, Sebastiá MT, Mengual J. 2013. Principal component analysis applied to remote sensing. Modelling in Science Education and Learning, 6: 83-89. doi:https://doi.org/10.4995/msel.2013.1905.

Green AA, Berman M, Switzer P, Craig MD. 1988. A transformation for ordering multispectral data in terms of image quality with implications for noise removal. IEEE Transactions on geoscience and remote sensing, 26(1): 65-74. doi:https://doi.org/10.1109/36.3001.

Guan H, Liu H, Meng X, Luo C, Bao Y, Ma Y, Yu Z, Zhang X. 2020. A quantitative monitoring method for determining Maize lodging in different growth stages. Remote Sensing, 12(19): 3149. doi:https://doi.org/10.3390/rs12193149.

Hyvärinen A, Oja E. 1997. A fast fixed-point algorithm for independent component analysis. Neural computation, 9(7): 1483-1492. doi:https://doi.org/10.1162/neco.1997.9.7.1483.

Ibarrola-Ulzurrun E, Marcello J, Gonzalo-Martin C. 2017. Assessment of component selection strategies in hyperspectral imagery. Entropy, 19(12): 666. doi:https://doi.org/10.3390/e19120666.

Javan F, Hasani Moghaddam H. 2017. Deforestation detection of Hyrcania forest using satellite imagery and Support Vector Machine (Case study: Rezvanshahr county). Forest Strategical Approchment Journal, 2(5): 1-13. https://www.magiran.com/paper/1706792. (In Persian).

Li X, Chen W, Cheng X, Liao Y, Chen G. 2017. Comparison and integration of feature reduction methods for land cover classification with RapidEye imagery. Multimedia Tools and Applications, 76(21): 23041-23057. doi:https://doi.org/10.1007/s11042-016-4311-4.

Luo G, Chen G, Tian L, Qin K, Qian S-E. 2016. Minimum noise fraction versus principal component analysis as a preprocessing step for hyperspectral imagery denoising. Canadian Journal of Remote Sensing, 42(2): 106-116. doi:https://doi.org/10.1080/07038992.2016.1160772.

Manly BF, Alberto JAN. 2016. Multivariate statistical methods: a primer. Chapman and Hall/CRC. 269 p. https://doi.org/10.1201/9781315382135.

Matkan AA, Nohegar A, Mirbagheri B, Torkchin N. 2014. Assessment relations of land use in heat islands using time series ASTER sensor data (Case study: Bandar Abbas city). Journal of RS and GIS for Natural Resources, 5(4): 1-14. http://girs.iaubushehr.ac.ir/m/article_516652.html?lang=en. (In Persian).

Nascimento JM, Dias JM. 2005. Vertex component analysis: A fast algorithm to unmix hyperspectral data. IEEE transactions on Geoscience and Remote Sensing, 43(4): 898-910. doi:https://doi.org/10.1109/TGRS.2005.844293.

Pohl C, Van Genderen J. 2016. Remote sensing image fusion: A practical guide. Crc Press. 288 p. https://doi.org/10.1201/9781315370101.

Pu R. 2017. Hyperspectral remote sensing: fundamentals and practices. CRC Press. 575 p.

Richards JA, Richards J. 2013. Remote sensing digital image analysis. Springer, Edition Number 5, XIX, 494 p. https://doi.org/10.1007/978-3-642-30062-2.

Sheffield C. 1985. Selecting Band Combinations from Multi Spectral Data. Photogrammetric Engineering and Remote Sensing, 58(6): 681-687. https://ci.nii.ac.jp/naid/80002491091.

Strîmbu VF, Strîmbu BM. 2015. A graph-based segmentation algorithm for tree crown extraction using airborne LiDAR data. ISPRS Journal of Photogrammetry and Remote Sensing, 104: 30-43. doi:https://doi.org/10.1016/j.isprsjprs.2015.01.018.

Vidhyavathi R. 2017. Principal component analysis (PCA) in medical image processing using digital imaging and communications in medicine (DICOM) medical images. International Journal of Pharma and Biol Sciences, 8(2): 598-606. doi:http://dx.doi.org/10.22376/ijpbs.2017.8.2.b.598-606.

Wang L, Zhang J, Liu P, Choo K-KR, Huang F. 2017. Spectral–spatial multi-feature-based deep learning for hyperspectral remote sensing image classification. Soft Computing, 21(1): 213-221. doi:https://doi.org/10.1007/s00500-016-2246-3.

Yang M-D, Huang K-H, Tsai H-P. 2020. Integrating MNF and HHT transformations into artificial neural networks for hyperspectral image classification. Remote Sensing, 12(14): 2327. doi:https://doi.org/10.3390/rs12142327.

Zhang X, Chen CH. 2002. New independent component analysis method using higher order statistics with application to remote sensing images. Optical Engineering, 41: 1717-1728. doi:https://doi.org/10.1117/1.1482722.

Zhang Y, Zhang J, Yang W. 2020. Quantifying Information Content in Multispectral Remote-Sensing Images Based on Image Transforms and Geostatistical Modelling. Remote Sensing, 12(5): 880. doi:https://doi.org/10.3390/rs12050880.