الفهرس 
Title : Enhanced Deep Learning Techniques for Spatial Big Data
contentsOnly 14 pages are availabe for public view
 |  | from | 159 |  |  |
| | | from | 159 | | |
Abstract
Deep learning (DL), an innovative version of neural network, becomes a hotspot
research topic in the Remote Sensing (RS) field. Recently, different deep learning
architectures have been developed to meet the challenges of spatial big data era. In this
thesis, a comprehensive investigation of deep learning approaches was introduced to
better handle spatial big data. Following is an outline of the three primary contributions
made in this thesis:
• Object Detection is an important and difficult challenge in remote sensing
imagery for civilian applications such as traffic monitoring, military applications, and
ship/Airplane detection. These applications are critical for decision-makers at this
time. The adaptive mask Region Based Convolutional Neural Networks (RCNN) was
utilized to detect multi-scale objects in optical remote sensing images. Experiments were
conducted to determine the efficacy of the following optimization methods: Adam, SGD,
RMSprop, Adadelta, hybrid SGD_Adam, and hybrid Adam_SGD in RS domain.
According to the obtained results, the average precision (AP) of the different
optimization methods were 90.8%, 87.7%, 87.3%, 48.6% and 91.2%, respectively. Using
SWATS (switch from Adam to SGD) in the phase of training. SWATS reduced
computation cost and time while achieving excellent accuracy. The proposed adaptive
MaskR-CNN outperformed other deep learning object detection methods in terms of
mean AP such as FRCN, YOLO, YOLO2, SSD, R_FCN were 95%, 76.4%, 66.7%, 79.7,
89.4%, 92.8%, respectively.