Provide a method for targets detection in satellite imagery using deep learning with remote sensing and GIS approach

Subject Areas : RS & GIS

nader biranvand ¹ , Mehdi Keykhaee ² , rooin mamani ³

1 - Emam Ali
2 - PhD Student in Remote Sensing and GIS, University of Tehran, Tehran, Iran
3 - Master of Civil Engineering, Shahid Beheshti University, Tehran, Iran

Received: 2021-05-26 Accepted : 2021-01-20 Published : 2021-05-22

Keywords: Remote sensing, deep learning, Convolution Neural Network, Automatic Target Detection,

Abstract :

Automatic detection of features in different areas according to various objectives, including urban management, military objectives, etc., are one of the most up-to-date and important applications of machine learning today. Combining the Global Geographic System (GIS) with images taken from satellite sensors and finally using deep learning methods, which is one of the main branches of machine learning, is a great help to the visible subject. Made the effects in the images using remote sensing science.. At the beginning of this research, the various layers in the proposed algorithm have been comprehensively presented and introduced, and then a new method has been presented in the simultaneous combination of CNN and pooling layers in the algorithm used, which finally It led to a significant reduction in network training time using comprehensive training data with high accuracy and at the same time high volume, which in the end, after entering the fully connected layer to extract and identify the desired goals with acceptable accuracy along with cost-effectiveness. Save time. In this research, using network training through training data, ships in satellite images are detected by creating a fully convoluted FCN network. In order to evaluate the performance and accuracy of the algorithm used in finding and detecting ships in satellite images, by applying this detection algorithm on several other satellite images, Precision, Recall and F1-Score evaluation criteria were used. The values were equal to 100%, 97.61% and 98.83%, respectively, which indicates the accuracy and reliability of the algorithm.