Deep Learning-Based Land Cover Classification for Satellite Images of Sri Lanka Using ConvNets

Abstract

Satellite imagery plays a vital role in remote sensing. People require expertise in interpreting satellite images to impend better and more effective decisions. With the advancement of technology and the availability of data, there is an increasing trend towards deep learning in the field of data analysis. We propose an image classification model based on deep learning to classify land cover in satellite images of Sri Lanka. In this research, we use satellite imagery data from Eurosat as our experimental data. Firstly, we categorize the data containing 12,000 RGB images into three main classes: urban, vegetation, and water. Secondly, we utilize image normalization as a pre-processing technique. We split 9,000 images as training and 3,000 images to validate the model. For the testing purpose, we create another dataset with 180 Sri Lankan satellite images. We use data augmentation and normalization as data prepossessing techniques. We configure our model with SGD optimizer and Categorical Crossentropy as the loss function. Training over 100 epochs with 16 batch size. This model gets overall accuracy of 65.33%. Model works fine with the Urban and the Water classes with the accuracy of over 95% for both the classes. All the vegetation images are classified as water. This is due to the spectral resolution of the satellite images that we use to train. Since we consider only red, green, and blue bands, it is hard to differentiate the vegetation and water as their reflectance rate are almost similar within the visible range of the electromagnetic spectrum.