Deep Convolutional Neural Network Models for Land Use and Land Cover Identification Using Dataset Created From LISS-IV Satellite Images

Abstract

Identification of land covers like crop-land, settlement, water-body and others from remote sensing images are useful for applications in the area of rural development, urban sprawl etc. In this paper we are addressing the task of identification of different land covers using remote sensed images which is further useful for image classification. Deep learning methods using Convolutional Neural Networks (CNN) for remote sensed or satellite image classification is gaining a strong foothold due to promising results. The most important characteristic of CNN-based methods is that prior feature extraction is not required which leads to good generalization capabilities. In this paper firstly we are presenting dataset prepared using multispectral, high-resolution images from LISS-IV sensor and another dataset of PAN images created using coarse-resolution images from Landsat-8 sensor. LISS-IV dataset is prepared for six commonly found land covers i.e. crop-land, water-body, bare-farm, road and settlement. Secondly we are proposing two patch-based Deep Convolutional Neural Networks (DCNN) models for prediction/identification of the land covers present in the image. Experiments conducted using the LISS-IV dataset has shown promising accuracies on both the DCNN models. Implementation of network is made efficient by harnessing graphics processing unit (GPU) power which reduces computation time. And finally, DCNN models are also evaluated for their performance using two similar publicly available benchmarked datasets, indicating that construction of models using described size of filters, number of filters and number of layers is suitable for multi-class remote sensing image patch prediction or identification.