| Title |
Classification of satellite images to create a map of the settlement area / |
| Authors |
Buinauskas, Laurynas ; Gadeikytė, Aušra ; Butkevičiūtė, Eglė |
| DOI |
10.15388/DAMSS.14.2023 |
| ISBN |
9786090709856 |
| Full Text |
|
| Is Part of |
DAMSS 2023:14th conference on data analysis methods for software systems, November 30 – December 2, 2023, Druskininkai, Lithuania.. Vilnius : Vilnius University press, 2023. p. 17.. ISBN 9786090709856 |
| Keywords [eng] |
settlement area ; convolutional neural networks ; satellite images classification ; machine learning |
| Abstract [eng] |
The growth of deep learning technologies has transformed image classification, especially by improving its ability to analyze satellite imagery for accurate identification of urban structures. However, satellite imagery classification of urban structures is often limited due to resolutions, building shape variances, overlapping, the complexity of the background, etc. This study aims to use advanced deep learning algorithms to automate the classification of satellite-taken land images, in order to facilitate the creation of detailed urban area maps. The study was conducted using object detection algorithms like AlexNet, VGGNet, InceptionNet, and ResNet. The main task was to achieve high model accuracy and efficiency. For this reason, the methodical data preparation strategy includes resizing and processing the original images. It should be noted, that in such a manner the dimensions of the obtained images are divisible by two, while also ensuring the representation of the original imagery without losing too much information. Concurrently, the generation of binary masks delineating building footprints is essential for data preparation. Various pre-processing strategies were explored to augment the model’s learning efficacy from the data. Classification of satellite images was done using the TensorFlow machine-learning framework. Metrics such as Intersection Over Union (IOU) and Pixel Accuracy were evaluated to ensure the accuracy of building segmentations in satellite images. The proposed classification system demonstrates the potential to efficiently identify buildings from other objects in each area that could be used for more precise urban planning and infrastructure development strategies. The presented system is capable of generating accurate urban maps and identifying unauthorized constructions, thereby contributing to enhanced infrastructure planning. |
| Published |
Vilnius : Vilnius University press, 2023 |
| Type |
Conference paper |
| Language |
English |
| Publication date |
2023 |
| CC license |
|
