A multi-model ensemble approach for quantifying sub-pixel land cover fractions in the urban environments

Abstract

ABSTRACTLand cover information is essential for sustainable management of the environment in urban areas. Satellite images have increasingly been used to extract such information, yet the accuracy has been challenged by the spectral and spatial heterogeneity of urban land covers. This paper presents a framework to develop a more skilful and reliable model for estimating land cover fractions using a multi-model ensemble technique, named Bayesian Model Averaging (BMA). The BMA is a statistical technique that combines the estimates of different models using Bayesian probability theory. In the BMA, each individual estimate is assigned a weight that is optimised in such a way that the likelihood of an individual estimate given the observation is maximised. In this study, three methods, viz. Multi-layer Perceptron (MLP), Pre-screened and Normalised Multiple Endmember Spectral Mixture Analysis (PNMESMA) and Support Vector Regression (SVR) have been used to develop an Ensemble Model (EM). We used a cluster-based approach for applying the BMA to utilise the diverse advantages in individual models. First, the image pixels were separated into three clusters by applying Normalised Difference Vegetation Index (NDVI) thresholds. Second, an ensemble of models for each cluster was derived using the BMA, and these ensembles were finally combined to derive the final output. The EM was tested in a heterogeneous urban area, viz. South Delhi, India, using two multi-spectral images, including Landsat Enhanced Thematic Mapper Plus (ETM+) and Advanced Spaceborne Thermal Emission Reflectance Radiometer (ASTER). The modelled land cover fractions were compared with the reference land cover fractions derived from a high-resolution (approximately 1 m) panchromatic image of the OrbitView-3 satellite. The accuracy assessment revealed that the EM estimates more accurate and reliable land cover fractions than the individual models on both the images. The performance of the EM in terms of Root Mean Square Error (RMSE), bias and kappa coefficient (k) is generally superior to that of the best of the individual models. These findings can help improve the accuracy of land cover fractions in heterogeneous landscapes by combining the outputs of various diverse models.