This research employs fully convolutional neural networks, followed by the transfer learning method to extract buildings. The model was developed by utilizing layers of down sampling and upsampling. Two convolution layers and a ReLU activation function make up this model. To minimize overfitting the dropout layer is included. The outputs establish that the FCN model adequately predicts the pixels that correspond to buildings, but it also incorrectly predicts many non-building pixels as building pixels. The methodology of transfer learning using U-Net and the pre-trained model is utilized to improve the precision. The segmentation model library is employed, which provides access to about 25 encoders that can be used with the U-Net model. To improve outcomes, we used three alternative encoders. The study ${}^{\prime}\mathrm{s}$ findings show that the model performs better with Inception-V3 and U-Net than the other two encoders. The accuracy of the fully convolutional neural networks model was 89.86 while the accuracy of the Inception-V3 and U-Net architecture was 96.39 percent.
Paper
Full text
Building Extraction from Remote Sensing Images Using Deep Learning and Transfer Learning
Semantic Scholar · Computer Science · 2022
Abstract
This research employs fully convolutional neural networks, followed by the transfer learning method to extract buildings. The model was developed by utilizing layers of down sampling and upsampling. Two convolution layers and a ReLU activation function make up this model. To minimize overfitting the dropout layer is included. The outputs establish that the FCN model adequately predicts the pixels that correspond to buildings, but it also incorrectly predicts many non-building pixels as building pixels. The methodology of transfer learning using U-Net and the pre-trained model is utilized to improve the precision. The segmentation model library is employed, which provides access to about 25 encoders that can be used with the U-Net model. To improve outcomes, we used three alternative encoders. The study ${}^{\prime}\mathrm{s}$ findings show that the model performs better with Inception-V3 and U-Net than the other two encoders. The accuracy of the fully convolutional neural networks model was 89.86 while the accuracy of the Inception-V3 and U-Net architecture was 96.39 percent.