Bias-Aware Data Quality Enhancement for Forest Fire Detection in AI-Based Remote Sensing

Bias, including preconceived notions embedded within the algorithm, selection in dataset collection, imbalances within the training data, impose severe restrictions on the algorithm's performance for accurate detection of forest fires. The variability of forest landscapes, environmental conditions across different geographic regions exacerbates these challenges. Considering the importance of tailored feature extraction, debiasing strategies in improving model performance for critical applications like forest fire detection, a novel approach using dual autoencoder architectures is proposed to independently learn fire-specific, nonfire-specific features. The proposed method effectively addresses visual biases in the data sets, utilizes superimposed image synthesis to create overlay images that enhance feature representation. The proposed integrated framework, validated on the publicly available Wildfire, DFire data sets, shows significant improvements in detection accuracy, robustness. The proposed approach outperforms existing methods with an accuracy of 80%, 76% on the Wildfire, DFire data sets, respectively, achieving lower false positive rates, false negative rates compared to the state-of-the-art methods.