Abstract:

Abstract With the development of Convolutional Neural Networks and increased processing power in recent years, the discipline of deep learning and machine learning has made significant advancements. One of the most important networks in the deep learning space is the Convolutional Neural Network. In computer vision and natural language processing, convolutional neural networks have achieved remarkable successes. Based on the land usage and land cover of the specific area, satellite images are useful to constantly monitor. Classifying satellite images using cutting-edge deep learning is one of the potential and difficult tasks of remote sensing. Three of the most popular Convolutional Network models viz., Custom Architecture, VGG16, and Resnet34, were utilised for classification in order to assess and investigate deep learning convolutional models utilising satellite data. The multispectral Sentinel 2 image with its 13 spectral bands served as the training data for these three models. The dataset of the study area was created manually and the featured images were classified into six classes. The accuracy for VGG16 was found to be 90.70% and that using Resnet34 and custom architecture was respectively 91.50% and 93.73%, thus demonstrating the fact that Custom architecture produces more accurate results than the other two transfer learning techniques.

Keywords: