Title:
TRANSFER LEARNING STRATEGY FOR SATELLITE IMAGE CLASSIFICATION USING DEEP CONVOLUTIONAL NEURAL NETWORK

Authors:
Abubakar Salihu Abba, Kabiru Ibrahim Musa, Abubakar Umar, Naziru Saleh, Hussaini Muhammad Khamis, Muhammad Kabir Dauda, Nigeria

Abstract:

Machine Learning is an essential research field as it has been utilized in remotely sense image classification. In image classification, data are made up of lots of samples characterized by many datasets. Hence high level of accuracy in terms of classification and training performance is such a big challenge. Machine learning techniques have broadly employed to build a substantial and accurate classification models. This paper proposes a new classification technique for remotely sense image classification, which is called Transfer Learning-Convolutional Neural Network (TL-CNN). TL-CNN is the introduction of transfer learning Pretrained on ResNet to convolutional neural network using Aerial Image Dataset (AID) with over 10000 images within 30 classes.  The remote sense image classification dataset consists of satellite images and not photographs, yet CNNs Pretrained on Remote Sense image has shown the ability to transfer to other image domains. This research work shows that, the proposed TL-CNN was able to improve the classification accuracy and returned 99.99% accuracy as against some of the existing algorithms which includes CNN with accuracy of 99.91%, SAE with accuracy of 93.98% and DBN with accuracy of 95.91%. The result also shows that, Using Pretrained CNNs as a starting point for fine-tuning on the ResNet image not only increases the training accuracy but also boosts the classification accuracy.