Comparison of deep learning methods for determining burnt forest areas with burnt area indices: The case of Hatay

In this study, the burnt areas and intensity of forest fires that occurred in the Samandağ region of Hatay between September 5-10, 2020, the Belen region of Hatay between October 9-10, 2020, and the Denizciler locality of Hatay between October 27-29, 2020, are mapped. It was analyzed using deep learning, remote sensing, and satellite data from Sentinel 2. With Sentinel 2 satellite photos of the research locations, an image dataset for deep learning was constructed. Then, using deep learning approaches, a deep learning model was developed, trained using the photos in the dataset, and successfully tested. Images from Sentinel 2 were used to produce the Normalized Fire Intensity (NBR) and Burnt Area Index for Sentinel 2 (BAIS2) indices using the results of a new deep learning model. Calculating the Difference Normalized Burning Intensity (dNBR) and Burnt Area Index for Difference Sentinel-2 (dBAIS2) values for the discrepancies between these indices before and after the fire allowed for categorization and determination of the fire area. The deep learning approach, burnt area indexes, and General Directorate of Forestry fire registration slips were compared, and it was established that the new deep learning model was more effective at locating burned forest areas than the indexes. In identifying the burnt forest areas, the new model has a proportionate accuracy of 98,36% in the Samandağ study region, 99,46% in the Belen research area, and 92,75% in the Denizciler study area. Key Words: Deep Learning; Sentinel 2; NBR-dNBR; BAIS2-dBAS2; Remote Sensing