Land Cover Classification from Satellite Data using Machine Learning Techniques

Abstract

This work attempts automatic land cover classification of different parts of India into forest, built-up, agricultural land and water bodies using temporal remote sensing data. Data from Agra district, Uttar Pradesh has been used to train different models - k-nearest neighbours, decision trees, support vector machines and convolutional neural networks. These models are then tested in Ahmedabad and Gandhinagar, Gujarat. Google Earth Engine has been used to obtain data from Landsat 8 satellite images. For the purpose of classification, Normalized Difference Vegetation Index (NDVI) values are calculated by masking all other light bands except near-infrared and red light bands. Temporal images with NDVI labels are fed as input to train the models and subsequently, the performance of these models is compared. A convolutional neural network based on the U-Net architecture is found to produce the most accurate results, improving upon traditional machine learning techniques. The models implemented can be used to produce land cover maps for any region, with good accuracy, which can then be used for various applications like natural resource management, urban expansion etc.