Deep learning instance segmentation framework for burnt area instances characterization

Abstract

The resemblance of burnt areas with other bright features undermines the certainty of wildfire detection. Bare surfaces and water reflection mislead and directly affect the detection rate. As of now, burnt area characterization and detection of resembling bright features are confined to conventional approaches (change detection, machine learning techniques, semantic segmentation). Consequently, the presented research article established an innovative deep learning instance segmentation model ahead of semantic segmentation. Transfer learning is employed on the ResNet-50/101 as the backbone. For burnt area detection and segmentation, the best performance with deep learning reported in the literature was 98%. The proposed technique was trained using variant regions (datasets) and evaluated precision based on IOU threshold, F1-Score, kappa, recall, missed & detection rate, with an overall accuracy of 98.5%. The research work provides the accurate groundwork for the hybrid qualitative and comparative quantitative analysis among classifiers (U-Net Classifier), capsule-based segmentation models (SegCaps, BA_EnCaps), semantic segmentation models (PSPNET, DeepLabV3) keeping the backbone (ResNet-50) and hyperparameters configuration identical. The suggested model indicated that the instance segmentation deep learning approach outperforms primitive techniques by presenting a greater detection rate and segmentation accuracy. The research inferred that compared to primitive approaches, integration of bright and resemble feature detection support burnt area characterization that localizes and characterizes each smallest fragmented overlapped burnt area instance (feature part).

Keywords

• Deep learning
• Instance segmentation
• Automatic burnt area extraction
• Bright surfaces