Vegetation cover change detection (2000–2020) in Raub District, Malaysia using supervised and unsupervised classification techniques

This study evaluates and compares four algorithms: maximum likelihood classification (MLC), minimum distance classifier (MDC), parallelepiped classifier (PPC), and k-means clustering, on their performance in detecting change of vegetation cover from multi-temporal Landsat images in Raub District, Malaysia in 2000–2020. The accuracy assessment is based on 150 stratified random points using overall accuracy (OA), user’s accuracy (UA), producer’s accuracy (PA), and kappa coefficient (KC). MLC resulted as the most accurate for all the years with OA ranging from 86.8 to 92.7% and kappa values between 0.79 and 0.86; besides it was found to be better for distinguishing some spectrally similar land covers such as oil palm (UA = 90%; PA = 90–100%) and rubber plantation (UA = 70–90%; PA = 100%). The k-means result has moderate accuracy with KC values ranging from 72.0 to 85.3% but this method mostly confused open grassy area with sparse forest area. MDC and PPC did not perform well; OA got as low as 47.6%. Moreover, high omission error was found in PPC when certain classes were absolutely omitted in a particular year. Dense forest decline was more than 27% (115,000–118,200 ha in 2000 to 83,500–86,100 ha by 2020), accompanied by about an 88% increase in areas under oil palm plantations. Open grassy areas doubled. The overall classification performances ranged best to poorest as: MLC > k-mean > MDC > PPC. These findings validate that MLC is the most dependable method for instituting policy-relevant monitoring over highly heterogeneous tropical landscapes while k-means delivers a rapid prelude and MDC/PPC are options with limited data or computational constraints.