SEMANTIC SEGMENATION OF LISS-4 SATELLITE IMAGERY FOR MAPPING BUILT-UP LAND AGGREGATES USING DEEP LEARNING

Hits: 31
Research areas:
Year:
2017
Type of Publication:
Article
Keywords:
convolutional neural network, segmentation, remote sensing, built-up land
Authors:
Pankaj Bodani, Kriti Rastogi Ujjwal Gupta
Abstract:
Mapping built-up land is important for city growth monitoring and building models for city growth forecasting. Aerial and satellite ortho-imagery is popularly used for this purpose. Knowledge-based, semi-automated approach for this task is challenging and requires significant human intervention and application of subjective expertise for iterative refinement of rule sets used for semantic segmentation. This is due to high heterogeneity in shape, density and composition of built-up land aggregates. This paper investigates a fully automated deep learning approach based on convolutional neural network for this task. This eliminates the need of human expertise for defining complex rule sets for semantic segmentation. Specifically, this paper discusses design, training and performance evaluation of a deep convolutional neural network based on SegNet architecture and presents analysis of design choices for the network. SegNet is a deep encoder-decoder architecture for semantic scene segmentation. The network is trained on multi-spectral LISS-4 satellite orthoimagery covering the central core and peripheral developing areas of a city. This paper also discusses algorithmic preprocessing and data augmentation techniques that resulted in improvement in accuracy and generalization.
Full text: 23.pdf [Bibtex]
You are here: Home ACRS ACRS Overview Proceedings semantic-segmenation-of-liss-4-satellite-imagery-for-mapping-built-up-land-aggregates-using-deep-learning