الفهرس | Only 14 pages are availabe for public view |
Abstract During the past few years, the number of earth observation satellites has increased, and the ability of these satellites to scan ne details and characteristics of the Earth surface has also increased. These details include the spatial, spectral, and temporal dimensions in their ne details. All this huge amount of data is scanned and recorded in data centers around the world. Much of this data is now accessible at a little cost or even oered for free. This huge amount of data has become impossible to process and analyze with classical computational ways. All this made the processing, analysis, and classication of satellite images an indispensable necessity for the real benet of satellites. Therefore, this thesis provides an integrated framework for classifying the dierent types of satellite images, using innovative methods to increase classication accuracy and increase the eciency of classi cation performance compared to other modern and public methods. In terms of spatial resolution, the proposed framework included classifying every high spatial-resolution image (less than 1 m) and classifying proportionally, low-resolution (greater than 10 m) images. In terms of spectral resolution, the proposed framework classied low spectral-resolution images with 3 traditional bands (red, green, and blue) and medium-resolution spectral images with 10 bands, as well as hyperspectral images with more than 100 bands. In terms of temporal resolution, we have used the classication of satellite images during a single season several times to help in the classication process. The proposed framework utilizes and customizes the state-of-the-art deep learning methods as well as traditional methods for comparison and evaluation. A deep review and study have been carried out for the dierent existing deep learning frameworks. These dierent frameworks exhibit dierent conceptions to the examined problem, including the use of many neuron layers, types of layers (convolution, pooling, etc.), and the type of connection between layers. As a customization process, we dealt with data preparation to meet our remote sensing context as well as optimization of learning parameters such as learning rate, loss functions, dropouts, etc. The proposed framework has been coded in Python as a leading open-source language and utilized the up-to-date scientic and learning libraries such as Tensor ow, Sci-kit-learn, and Numpy and we oer it to the scientic community to use it build upon xiv it. To succeed with deep learning methods, we had to have huge datasets for training, validation, and testing. In this regard, we used the available benchmark data that have been reported in the literature as well as our own remote sensing data for some real applications. This framework has been applied in many practical and experimental applications. It has been applied the proposed methods to classify the dierent types of agricultural cover in Fayoum Governorate in Egypt using images of the Sentinel 2 and Landsat 8 satellites. It has, also, been applied to classify dierent levels of urban areas in Greater Cairo through the Sentinel2 satellite imagery. In the object detection context, the framework has been applied to high spatial-resolution images to detect and localize objects like planes, cars, ships, etc. For spectrally rich images, a new method has been proposed to classify vegetation cover using hyperspectral images. All proposed methods have been conceptualized so that best performance, whether in terms of computational speed or classication accuracy, is to be achieved. To achieve this, the standard evaluation metrics have been used to evaluate methods in the proposed framework in dierent hardware and software environments (desktops, cloud computers, Windows operating systems, Linux operating systems) The results showed a remarkable superiority with the proposed framework methods. |