Geographia Technica, Vol 21(2), Special Issue: Artificial Intelligence Applications in Geography, 2026, pp. 158-187
DEEP LEARNING-BASED DETECTION OF AGRICULTURAL BURNED AREAS USING TRUE COLOR SATELLITE IMAGERY
Worawit SUPPAWIMUT 
, Ratchaphon SAMPHUTTHANONT
ABSTRACT: Chiang Mai Province, situated in northern Thailand, is experiencing persistent challenges regarding seasonal haze and particulate matter (PM2.5) contamination. The primary sources of these emissions are agricultural burning practices, particularly the combustion of rice straw and stubble in proximity to residential areas. This study aims to (1) assess agricultural burned areas in rice fields using the differenced Normalized Burn Ratio (dNBR) derived from Sentinel-2 imagery, and (2) develop a deep learning model to detect burned areas from true color satellite imagery (RGB) using the U-Net architecture. Sentinel-2 Level-2A images acquired before and after burning events in 2024 were used for analysis. The study focused on rice cultivation areas, while clouds, cloud shadows, and water bodies were removed using the Scene Classification Layer. Subsequently, the Normalized Burn Ratio (NBR) and the Differenced Normalized Burn Ratio (dNBR) were calculated to generate a burned area mask. The validation results indicate that a dNBR threshold of ≥ 0.1 is appropriate for the specific context of the study area. This mask was used as reference data for training the deep learning model. The U-Net model was trained using RGB image tiles and corresponding binary masks, with the final deep learning dataset consisting of 2,560 image tiles of 256 × 256 pixels after data augmentation. A spatial hold-out strategy was applied to prevent spatial data leakage between training, validation and test datasets. The results show that U-Net outperformed DeepLabv3+, achieving an Intersection over Union (IoU) of 0.9292 and an F1-score of 0.9633, indicating strong capability in detecting small burned patches. Spatial analysis revealed that the burned areas accounted for 9.952% of the total image area and 18.937% of the total cultivated land. These findings suggest that integrating true-color satellite imagery with deep learning techniques provides significant potential for monitoring agricultural burning and supporting local strategies to mitigate PM2.5 pollution.
Keywords: Deep Learning, Agricultural Burned Area, Normalized Burn Ratio, Satellite Imagery
Full article here