13 Aralık 2019, Cuma

Deep Learning

In recent years, many domains of science, business and government use deep learn- ing technology in the world. Deep learning is the field of artificial intelligence and computers can learn many things by using large amounts of data without human involvement. The algorithms of deep learning inspired by the human brain, hu- man brain learns from experience, deep learning algorithm would perform a task repeatedly and each time evolves the outcome by using given data

Deep learning is characterized by neural networks (NNs) and it usually has more than two hidden layers, so they are called ’deep’. Basic neural network is shown in the Figure 1 and each connection between neurons is associated with a weight such as wij , wjk , wkl. Each neuron performs dot product with its input matrices that contain all pixels of the images and its weight matrices. Then neuron sums all dot product results and adds bias. After that, each neuron applies the non-linearity which is also called activation function. For our case activation function is ReLu which will be explained. Figure 1.3 shows the mathematical model of neuron.

Neural Networks: They are organized in layers with set of nodes 

Deep Neural Networks make use of feature representations which are learned from data instead of handcrafting features which are designed domain-specific knowledge

Deep learning methods are used in target recognition and it can be divided into two main categories: supervised and unsupervised methods. Unsupervised method learns features from the input data but it does not know any correlated la- bel or information. Supervised method has prior knowledge, uses given sample data to make best approximates that between input and output data. Supervised method or learning is also divided into two categories that are classification and regression. Classification predicts the class the data belongs to and regression predicts the numeric value based on the observed data which is trained data

Deep learning has won remarkable success and its neural networks are powerful learning machines. It generalizes to different domains, for example, semantic segmentation model performs classifying pixels, the same model that is trained on different data can be used to predict oil fault lines and detect cars etc. High resolution satellite images are one of the resources and they have been an active research topic in the last decades. Increasing availability of data and computational resources provide the use of deep learning in remote sensing.

Mathematical Model of Neuron

Deep learning is being used to solve problems in the field of urban planning and computational sustainability. For instance, popular deep neural network which is convolutional neural network (CNN) has been used to estimate spatial distribution of poverty. CNN, which will be explained in next chapters, shows that it is effective at the problem of remote sensing image scenes classification


  1. J. Zhang, T.; Huang, X.; Wen, D.; Li,, “Urban building density estimation from high-resolution imagery using multiple features and support vector re- gression,” Ieee Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 10, no. 7, pp. 3265–3280, 2017.
  2. [2]  Y. Zhou, C. Lin, S. Wang, W. Liu, and Y. Tian, “Estimation of building density with the integrated use of gf-1 pms and radarsat-2 data,” Remote Sensing, vol. 8, no. 11, p. 969, 2016.
  3. [3]  B. Yu, H. Liu, J. Wu, Y. Hu, and L. Zhang, “Automated derivation of urban building density information using airborne LiDAR data and object-based method,” Landscape and Urban Planning, vol. 98, no. 3-4, pp. 210–219, 2010.
  4. [4]  Q. Wu, R. Chen, H. Sun, and Y. Cao, “Urban building density detection using high resolution SAR imagery,” in 2011 Joint Urban Remote Sensing Event, JURSE 2011 – Proceedings, 2011, pp. 45–48.
  5. [5]  X. Z. Pan, Q. G. Zhao, J. Chen, Y. Liang, and B. Sun, “Analyzing the variation of building density using high spatial resolution satellite images: The example of Shanghai City,” Sensors, vol. 8, no. 4, pp. 2541–2550, 2008.
  6. [6]  S. C. Liew, “Principles of Remote Sensing – Centre for Remote Imaging, Sens- ing and Processing, CRISP,” Interpreting Optical Remote Sensing Images, pp. 2–6, 2001