Identifying Settlements Using SVM and U-Net
| Autor: | Abhra Singh, Guneet Mutreja, Rohit Singh, Sandeep Kumar, Divyansh Jha |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
Training set
Computer science Process (engineering) business.industry Deep learning Training (meteorology) 02 engineering and technology Land cover 010501 environmental sciences Machine learning computer.software_genre 01 natural sciences Class (biology) Data modeling Support vector machine 0202 electrical engineering electronic engineering information engineering 020201 artificial intelligence & image processing Artificial intelligence business computer 0105 earth and related environmental sciences |
| Zdroj: | IGARSS |
| DOI: | 10.1109/igarss39084.2020.9324702 |
| Popis: | Deep learning (DL) has shown promising results, especially in the area of computer vision, making it suitable for remote-sensing image analysis tasks such as land cover classification. However, the increased accuracy comes at a cost of increased training data requirements. This paper showcases our approach to address this shortcoming by using machine learning to bootstrap the process. To achieve this aim, supervised classification has been carried out on Landsat 8 imagery using Support Vector Machines (SVM). Out of the five classes segmented using traditional supervised classification (SVM), urban class is taken into consideration for training the DL model. The dataset is manually cleaned before feeding it into the DL model for training. The model achieves an accuracy of 0.92 at identifying settlements after 1,100 training iterations. This approach saved time required to prepare training data for the DL model, while producing highly accurate results. Such a process can reap the benefits of DL for geospatial mapping, while reducing the requirement to start with copious amounts of labelled training data. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|