Machine Learning-Assisted Denoising of Raman Spectral Remote Sensing Data for Improved Land Use Mapping
Revue Internationale de Géomatique, (2025), Vol. 34, No. 1, pp. 415-432
Fawad Salam Khan
a
,
Noman Hasany
b
,
Sheikh Kamran Abid
c
,
Muhammad Khurram
d
,
Jerome Gacu
e,f,g
,
Cris Edward Monjardin
h
,
Kevin Lawrence M. De Jesus
g
a Campus Cité Scientifique, University of Lille, Villeneuve-d’Ascq, Lille, 59650, Hauts-de-France, France
b College of Computer Science and Information Systems, Institute of Business Management, Korangi Creek, Karachi, 75270, Pakistan
c Faculty of Technology Management and Business (FPTP), Universiti Tun Hussein Onn Malaysia, Batu Pahat, 86400, Johor, Malaysia
d Electronic Engineering Department, University of Technology and Applied Sciences, Nizwa, 611, Sultanate of Oman
e Civil Engineering Department, College of Engineering and Technology, Romblon State University, Odiongan, Romblon, 5505, Philippines
f Disaster Prevention Research Institute, Kyoto University, Gokasho, Uji, Kyoto, 611-0011, Japan
g Department of Civil Engineering, College of Engineering, FEU Institute of Technology, P. Paredes St., Sampaloc, Manila, 1015, Philippines
h School of Civil, Environmental and Geological Engineering, Mapua University, Manila, 1002, Philippines
Abstract: Noise present in remote sensing data creates obstacles to proper land use and land cover (LULC) classification methods. The paper evaluates machine learning (ML) denoising methods that adapt Raman spectroscopy’s spectral techniques to optimise remote sensing spectra for land-use/land-cover (LULC) mapping. A basic Raman spectroscopy model demonstrates that Savitzky-Golay (SG) filtering, Wavelet denoising, and basic 1D Convolutional Autoencoder have different effects on synthetic spectral features relevant to LULC classification. Savitzky-Golay filtering yielded the most efficient results, increasing classification accuracy from 0.71 (noisy) to 1.00 (denoised), resulting in perfect classification with zero errors and enhancing the Precision-Recall curve, as Area Under the Precision-Recall Curve (AUC-PR) transformed from 0.84 to 1.00. The study examined wavelet denoising in conjunction with a 1D Convolutional Autoencoder, assessing the noise reduction capability through visual evaluation. Based on Raman-based spectral analysis, a traditional method complemented with machine learning denoising provides promising fields for feature identification in remote sensing images, thereby improving the quality of LULC-related mapping outcomes.