Publications

Verma, P; Rajitha, B (2024). Weighted ensemble approach for smoke-like scene classification in remote sensing images. SIGNAL IMAGE AND VIDEO PROCESSING, 18(10), 7359-7367.

Abstract
Damage to the environment or the ecosystem because of natural lightning or human-induced circumstances can lead to losses that reverberate across decades. Harnessing satellite images can be an imperative change that can help avoid relentless and escalating fires. Continuous and comprehensive coverage by remote sensing enables the detection of small plumes of fire or smoke that can emerge as wildfire if not detected early. Mainly, research has been focused on only limited predefined classes. To overcome this limitation the dataset provided by the Moderate Resolution Imaging Spectroradiometer (MODIS), referred as USTC_SmokeRS is utilized in this study. This dataset includes smoke and situations that resemble smoke, such as clouds, dust, and haze. USTC- SmokeRS dataset helps in developing a system that can adapt across diverse geographical regions. This article employs three different pre-trained models: MobilenetV2, InceptionV3, and DenseNet201, which are optimized and are used for the smoke classification task. Along with optimization of the models, an ensemble algorithm is also proposed to combine the capabilities of the three individual models. The proposed ensemble model achieved the highest accuracy of 95.73% on this dataset. Further, the Grad-CAM technique is adapted for localization task.

DOI:
10.1007/s11760-024-03399-4

ISSN:
1863-1711