- Title
- Developing logit calibration model for wildfire smoke characterization based on sentinel-2 multispectral data and machine learning techniques
- Creator
- Sali, Athule
- Subject
- Wildfires -- Prevention and control -- Contracting out
- Subject
- Smoke plumes
- Subject
- Remote-sensing images
- Date
- 2023-07
- Type
- Master's theses
- Type
- text
- Identifier
- http://hdl.handle.net/10353/28467
- Identifier
- vital:74338
- Description
- Wildfires are complicated incidents that arise as results of both natural causes and anthropological activities. They have long been regarded as the most devastating phenomena globally. Wildfires are considered a powerful natural factor which has detrimental effect on the global environment. This study was aimed at developing logit calibration models for wildfire smoke prediction based on Sentinel-2 multispectral data and machine learning techniques. Remotely sensed data, in the form of the Sentinel-2 imagery, was used as the base from which wildfire smoke plumes were spectrally characterized and distinguished from clouds and flame using endmember selection. The Smoke Detection Index (SDI) was generated to detect the relative abundance of smoke from the imagery. The Cloud Detection Index (CDI) was also generated from Sentinel-2 imagery to detect the relative abundance of clouds. The bi-level thresholding technique was also used to characterize wildfire smoke from the imagery. The logit models were developed through multilayer perceptron (MLP) neural network to predict wildfire smoke plumes. The Relative Operator Characteristic - Area Under the Curve (ROC-AUC) metrics was used to evaluate the logit models performance. The spectral signature pattern from endmembers revealed that wildfire smoke behaves different across Sentinel-2 multispectral channels with shortwave 1 channel (SWIR-1) exhibiting the highest radiance value. The signature patterns from endmember selection also revealed the distinctive spectral characterization of smoke from those of clouds. The findings showed that whilst smoke exhibited high radiance value on SWIR-1 channel, clouds exhibited high radiance value in the near infrared (NIR), signifying that smoke and cloud are spectrally separatable in the NIR. The smoke-containing pixels from bi-level thresholding were characterized by SDI values that ranged between 0.089 and 0.561. Suggesting that pixels associated with wildfire smoke are limited to this range of values. The logit models developed showed that smoke is predicted in SWIR-2. The ROC-AUC value obtained by this model was 0.77. The Implications emerging from the ROC-AUC results revealed that MLP model employed on the SWIR-2 band present a viable and accurate prediction of wildfire smoke plume. The findings of this study suggest that wildfire smoke is efficiently predicted at the shortwave channels of the electromagnetic spectrum. The wildfire smoke can be spectrally distinguished from cloud in the near infrared channel.
- Description
- Thesis (MSc) -- Faculty of Science and Agriculture, 2023
- Format
- computer
- Format
- online resource
- Format
- application/pdf
- Format
- 1 online resource (146 leaves)
- Format
- Publisher
- University of Fort Hare
- Publisher
- Faculty of Science and Agriculture
- Language
- English
- Rights
- University of Fort Hare
- Rights
- All Rights Reserved
- Rights
- Open Access
- Hits: 298
- Visitors: 298
- Downloads: 59
Collections
UFH Department of GIS & Remote Sensing
| Thumbnail | File | Description | Size | Format | |||
|---|---|---|---|---|---|---|---|
| View Details | SOURCE1 | THESIS_SALI_Final.pdf | 47 MB | Adobe Acrobat PDF | View Details |