Application of principal component analysis to remote sensing data for deforestation monitoring

Abstract

Principal Component Analysis (PCA) is applied to Sentinel 2 Multi-Spectral Instrument (MSI) imagery to investigate its ability to detect deforestation through direct analysis of resulting Singular Values and Principal Component loading matrices. Initial work aims to compare deforestation detection in small areas across North-East Rondonia, Brazil with previous deforestation studies. Subsequently, a deforestation analysis of the Sentinel 2 MSI using PCA in Sub-Saharan Africa is presented. Standardised PCA is applied through the Singular Value Decomposition (SVD) of channel standardised, resampled input imagery. First, cropped sub-areas of input imagery are considered, termed local PCA. Local PCA is applied to images separately (separate rotation) and to two-year image composites (merged rotation). Both approaches were found to detect forest cover changes, with separate rotation allowing for the generation of time-series data. The change detection resolution of both approaches is relatively low; being able to detect only if change has occurred in the general area and not the exact location of greatest change. In order to improve change detection resolution and identify the sub-areas of greatest change, separate rotation PCA is applied on a pixel scale. A simple statistical threshold is used to implement bi-temporal change detection, demonstrating the ability of this approach to detect forest cover changes at a higher resolution. Lessons learned from the separate rotation local PCA technique is used to analyse forest cover changes in an area in Sub-Saharan Africa.