A Lightweight CNN Architecture for Land Classification on Satellite Images

Land cover classification using satellite images is an important tool in the study of terrestrial resources. Satellite based information is presently available as huge sets of high resolution images from a large number of satellites like Sentinel, Landsat-8, etc. Land cover classification from these images is a difficult task because of very large sized data and high variation types. Deep Neural Networks can play a vital role in this regard and can perform classification on these large sized data. Related works in this field have used lighter models and included a large number of handcrafted parameters which requires domain knowledge on the subject. It is realised that most models are too shallow for such a complicated image. In this paper, a deeper Convolutional Neural Network (CNN) model without any satellite image specific parameters is proposed. On SAT4 and SAT6 images, our 13-layered network has achieved better accuracy upto 99.84% and 99.47% which is state-of-the-art. It is still called lightweight model because most models in Artificial Intelligence(AI)-CNN are much deeper and larger than ours.