Fuzzy machine learning model to detect transition building footprints using multi-sensor multi-temporal images

Abstract

The research study proposes a fuzzy machine learning based approach for the detection of transitioning building footprints in urban area using temporal high and medium resolution datasets acquired by multi-sensors. Remote sensing dataset from Planetscope, Sentinel-2A/2B and Landsat-8 Operational Land Imager were used to generate the multi sensor based temporal indices database. The class-based sensor independent – normalized difference vegetation index (CBSI-NDVI) was used to reduce spectral dimensionality and to generate temporal database, which was given as an input in MPCM classifier with two training approaches, namely, mean training approach and Individual Sample as Mean (ISM) training approach. The training was performed using single and multiple phases of change in urban area in different time periods. The results were validated using high resolution Google Earth imagery as well as ground observations at five different testing sites. The kappa coefficient was calculated as 0.78 for the fuzzy MPCM with ISM training approach. Also, the MMD and variance were calculated to be 0.006 and 0.002 respectively. The study concluded that the MPCM classifier with ISM training approach could address the heterogeneity in urban area and provides classified images considering each transition trend as mean input for classification, thus improving the classification accuracy.