Deep Learning techniques for Geospatial Data Analysis

Abstract

Geospatial data increasingly drives applications across various domains, such as newlineurban planning, civil infrastructure enhancement, environmental monitoring, and newlinedisaster management. The advent of deep learning techniques, coupled with the newlineavailability of public datasets, has facilitated the development of innovative newlineapplications in these areas. newlineThe thesis addresses four distinct problems, namely (I) Mangrove mapping, (II) newlineLand Use Land Cover mapping, (III) Landslide prediction, and (IV) Black Spot newlineclassification. These problems were chosen to cater to the specific needs urban newlineplanners face in the Kokan region. By applying cutting-edge data analysis newlinetechniques, the research aims to address these local issues effectively. newlineThe approach adopted in this thesis involves selecting publicly available datasets newlineand pre-trained models, which are then fine-tuned using prepared datasets to newlinecapture local variations. This process enables the creation of task-specific data newlinemodels tailored to the Kokan region s unique characteristics. Deep learning newlinetechniques take centre stage improving the accuracy of the developed models. newlineThe main contributions are: (I) A comprehensive dataset that captures the local newlinevariations relevant to the chosen tasks. This dataset forms the foundation for the newlinesubsequent analysis and model development. (II) The development of fine-tuned newlinemachine-learning models specifically designed to tackle the selected tasks. These newlinemodels are optimized to address the specific challenges posed by each problem and newlineprovide efficient and accurate solutions. (III) The provision of a thorough evaluation newlineof the developed models. This evaluation ensures the practical applicability of the newlineproposed solutions. Furthermore, this thesis lays the groundwork for future research newlineand development of data-driven applications in geospatial domains. newline newline