TENSOR FLOW BASED FINE-TUNED CONVOLUTIONAL NEURAL NETWORK MODEL FOR SCENE CLASSIFICATION OF AERIAL IMAGES

Abstract

Scene classification of aerial images has received growing attention from the research community in recent 
years. Conventional classification algorithm such as k-means clustering, Support Vector Machine (SVM), 
Decision Tree, Expectation Maximization (EM algorithm), k-nearest neighors (k-NN), Ada Boost, Navie Bayes, 
Artificial Neural Networks (ANN) etc., use spectrum, texture and profile based feature for classification. 
Extracting such features in high resolution remote sensing images creates challenges to object classification. So 
we have to introduce the convolutional neural network (CNN) for classifying the remote sensing images. The 
architecture consist of three convolution layers of 64 filters with kernel size of 5x5, pooling size of 2x2 and fully 
connected layer 1024 and 512 respectively.  In the experiments, 8,000 remote sensing images has been collected 
from Pattern Net dataset for the performance assessment. The experimental results show the higher accuracy for 
the remote sensing image classification.