Comparative Assessment of Machine Learning Algorithms for Impervious Surface Mapping Using Sentinel-2 Imagery on Google Earth Engine, with a Neutrosophic Uncertainty Assessment: A Case Study of Aligarh District, India

Abstract

Abstract Urban expansion continuously transforms Earth's surface by replacing permeable soils with impervious cover, altering hydrological cycles and urban microclimate. Accurate delineation of impervious surfaces is therefore critical for understanding urban growth and its environmental consequences. Google Earth Engine (GEE) is an advanced cloud-based platform for analysing large volumes of geospatial data. This study assesses the effectiveness of four supervised machine learning classifiers: Random Forest (RF), Support Vector Machine (SVM), Gradient Tree Boosted (GTB), and Classification and Regression Tree (CART) for the extraction of impervious surfaces within the Google Earth Engine (GEE) framework. Sentinel-2 Level-2A imagery acquired from Jan to March 2023 was cloud-masked to generate a cloud-free composite for Aligarh District, India. The bands used were B2 (blue), B3 (green), B4 (red), and B8 (near infrared). A total of 4,520 points were manually labelled into two classes impervious and pervious with 3,262 points used for training and 1,258 for validation. Of the 1,258 validation points, 1,092 survived sampling after nodata masking and were used for accuracy assessment; because each classifier was exported as an independent Google Earth Engine task, GTB and CART were evaluated on 1,091 of these 1,092 points, a variation that does not affect the reported accuracies.The neutrosophic-weighted analysis reported below is computed on the 1,083-point subset for which all four classifier predictions could be matched. RF achieved the highest overall accuracy (OA = 97.89%, κ = 0.950), followed closely by GTB (97.71%, κ = 0.946) and CART (95.60%, κ = 0.896). SVM performed lower (91.21%, κ = 0.777), with the gap attributable to its sensitivity to overlapping spectral signatures in mixed and transitional pixels. The findings indicate that ensemble-based classifiers more effectively capture urban spectral variability, and they also demonstrate that GEE is a rapid and reproducible method for mapping urban surfaces on a regional scale. Because a single overall-accuracy figure cannot show where classifications are unreliable, a neutrosophic set-theoretic framework was additionally applied to the four classifier outputs, deriving a truth-indeterminacy-falsity (T, I, F) triplet for every pixel from inter-classifier agreement and producing a spatially explicit indeterminacy map of the study area. The methodology employed in this study can be adapted for ongoing urban monitoring and policy-oriented environmental assessments.