An Effective Forest Fire Prediction Based on Contrastive Vision Transformer with Pool Former

Since years, forest fires are a significant concern globally, causing damage to ecosystems and human settlements. For effective prevention and mitigation predicting forest fires is important, different factors contribute to forest fires which includes human activities, topography, climate change and weather conditions. Traditional approaches for predicting forest fires had faced significant challenges which include limited validation of prediction models and lack of standardization. Therefore, this paper proposes Contrastive Vision Transformer (CViT)-Pool Former for predicting forest fire which is employed on Forest Fire Big Data dataset which enabled better comparability. This dataset is preprocessed by using data augmentation, noise reduction and data enhancement techniques. Then, the features are extracted by using CViT which extracted optimal features based on weather conditions and captured contextual relationships by maximizing mutual information. After that, Pool Former light weight attention module is used for predicting fires which gave more accurate predictions by enabling targeted prevention and mitigation, minimized unwanted deployments and resource waste, boosted prediction accuracy by token fusion and enhanced representation learning. The proposed CViT-Pool Former method established better results including accuracy (92.8%), precision (89.4%), recall (91.4%) and F1-score (90.2%) when compared with existing Region Proposed Convolutional Neural Network (R-CNN).