Awesome

MUnet-LUC

Land Use Classification with mUnet by <a href = "https://github.com/abhi170599">Abhishek </a> and <a href="https://github.com/sayonpalit">Sayon</a>

<h2>Description</h2> <p>We aimed at developing a deep learning Pipeline to classify land use types from satellite images.<br> Land Use can be classified into following classes-<br> <ol> <li>Buildings</li> <li>Trees</li> <li>Crops</li> <li>Roads & Tracks</li> <li>Water</li> <li>Empty Fields</li> </ol> </p> <h2>Model Overview</h2> <p>We used the <a href= "http://www.insticc.org/Primoris/Resources/PaperPdf.ashx?idPaper=73706">paper </a>by Lakshya Garg et al to implement their proposed Modified UNet Architecture for Land Use Classification of Satellite Imagery</p> <br> <h3>Modified UNet Architecture</h2> <img src="https://raw.githubusercontent.com/abhi170599/MUnet-LUC/698627329b6f68000861a1140a680f225e415baa/Screen%20Shot%202020-05-20%20at%205.05.49%20PM.png" border = "5"> <h2>Results and Inference</h2> <img src = "https://github.com/abhi170599/MUnet-LUC/blob/master/Screen%20Shot%202020-05-19%20at%208.55.34%20PM.png"> <img src = "https://github.com/abhi170599/MUnet-LUC/blob/master/Screen%20Shot%202020-05-19%20at%208.53.52%20PM.png"> <p>Model was trained on <a href = "colab.research.google.com"> Colab's</a> <b>12GB NVIDIA Tesla K80 GPU</b> for 150 epochs <br> with training accuracy of 80.037%</p> <img src = "https://github.com/abhi170599/MUnet-LUC/blob/master/Screen%20Shot%202020-05-20%20at%208.09.30%20AM.png"> <table> <tr> <th>Model</th> <th>$Parameters</th> <th>Accuracy</th> </tr> <tr> <td>AlexNet</td> <td>61,000,000</td> <td>78.234%</td> </tr> <tr> <td>UNet</td> <td>31,379,205</td> <td>62.1077%</td> </tr> <tr> <td>mUnet</td> <td>31,105,669</td> <td>80.897%</td> </tr> </table> <h2>Applications</h2> <li>Smart City Planning (Searching Construction Space, Monitoring Vegetation etc.)</li> <li>Defence Applications</li> <li>Natural Resource Monitoring and Management ✓Disaster Management</li>