Multi-scale feature fusion segmentation and lightweight ensemble deep learning approach for urban slum classification using satellite images

Abstract

ABSTRACT Although there are some drawbacks to satellite imaging, such as the difficulty in identifying slums in highly populated metropolitan areas, it is nonetheless a useful tool for preliminary screening and can support on-the-ground and other data sources. Hence, we proposed the MSFF-Seg (Multi-Scale Feature Fusion-based Segmentation) model for urban slum classification using satellite images. Hence we proposed the MSFF-Seg model for urban slum classification using satellite images. Initially, three types of corrections are performed in preprocessing such as geometric correction which is corrected by Enhanced Scale Invariant Feature Transform (ESIFT), radiometric correction which is corrected by Ground-based Relative Radiometric Calibration (GRRC), and atmospheric correction which is corrected by homomorphic filter. After completion of preprocessing phase, segmentation is initiated by Tweak Mask RCNN which extracts the features in multilevel for segmentation that increases the segmentation accuracy. The segmented images are fed into an ensemble classifier for urban slum identification. Here, three types of deep learning are used for ensemble learning namely Schematic Convolutional Neural Network (SCNN), Shallow MobileNet, and Lite ShuffleNet. They extract multiple features from the segmented region, based on the extracted features. LGBM performed urban slum classification which classifies the images into two classes such as slums and non-slums. Finally, the performance of this research is evaluated based on several performance metrics like accuracy (98%), precision (96%), recall (94%), and F1 score (97%), which proved that the proposed work achieved superior performance in urban slum classification.