Deep Learning based Detection of Water Bodies using Satellite Images

A lot of ongoing developments are going on in the field of Deep Learning and then further in Image Processing. This paper aims to focus on the processing done using Convolutional Neural Networks to obtain the images having water pixels classified appropriately. The identification of water bodies and the knowledge about the geography of those regions is crucial to a lot of activities, it helps in further planning the developments in that region and in emergency operations too i.e. in rescue operations. The images are basically obtained through remote methods or by using low flying drones that capture them. However, in addition to the cost, following issues must be considered: remote satellite images may not trace sudden changes over a particular latitude and longitude while the drone may take a lot of time to capture all the details. The whole objective of this research is to find the locations of water bodies using the data available through images and then finding the area or region over which they are spread. The satellite images from Sentinel-2 have been used and the shape files too have been obtained in order to map the initial training data for the model. The mapped data is then stored in the form of preprocessed result and the model is trained further using the the preprocessed data. The time taken for the processing of the input images and the shapefiles depends highly on the machine being used, low end machines might crash while opening a shapefile because the size of the shapefile might go upto 1.5 GigaBytes. Eventually the whole process resulted into a trained classifier with an accuracy greater than 90% and with 70% precision.