DETECTING AND MEASURING ENVIRONMENTAL DISASTERS BASED ON IMAGE SEGMENTATION DEEP LEARNING TECHNIQUE

Abstract

It’s essential for any environmental monitoring system to detect and measure any sources of smoke or aerosol clouds that can pollute the air or predict environmental disasters like massive forest fires. To detect and monitor smoke in the air, aerial imagery is a very effective method, especially for large areas. The presence of aerosol and particles can affect transparency and spectral properties of the atmosphere, for this reason we can detect smoke in aerial images visually. Based on these properties, in this paper, we evaluate the possibility of using automated models for aerial image analysis and detection, fragmentation, and measurement of smoke clouds in image data collected using deep learning neural networks. We have built pixel-by-pixel labeled datasets in large numbers and improve and train our segmentation models that derived from Unet neural network architecture. The test models were evaluated by IoU (Intersection Over Union) measurement and false alarm rate. The test results demonstrate deep learning models that enable reliable and efficient detection of smoke in environmental and security applications.