SatNet: A Lightweight Satellite Image Classification Model Using Deep Convolutional Neural Network

Abstract

Deep Learning approach is essential for classifying remotely sensed images, identifying objects in satellite images, and using those results for numerous purposes. In light of the large number of datasets, image classification requires high accuracy and training performance, making the approach crucial to developing reliable and accurate classification models. As the field of satellite imaging has advanced, researchers are more inclined to rely on techniques for the automatic and efficient classification of satellite images. However, many well-known networks need a lot of parameters and computing power, which makes them challenging to use for classification tasks. A Lightweight Deep Convolutional Neural Network (Deep CNN) called SatNet is proposed for satellite image categorization which uses less computing power and is an efficient model for multiclass classification problems. Five transfer learning models - Xception, InceptionV3, VGG19, ResNet50, and DenseNet201 were taken into consideration in order to make comparisons. The Satellite Image Classification Dataset-Rsi-cb256is used by the model during testing. It effectively outperforms other measures and surpasses ResNet50 and VGG19 outputs, with an accuracy rate of 99.15%. A quicker and easier method of classifying satellite image data is possible with the proposed Deep CNN model.