LEO Satellite Oriented Wildfire Detection Model Using Deep Neural Networks: A Transfer Learning Based Approach

Wildfires are a recurring concern in several regions around the globe, often leading to catastrophic outcomes. The increasing accessibility of small satellites operating in Low Earth Orbit (LEO) has led to the potential for enhancing the speed of wildfire detection b y utilizing constellations of m any miniature satellites in the areas of interest. This research presents a system model for detecting and reporting wildfires to fire stations using LEO satellites. In this paper, we propose a transfer learning-based lightweight deep learning wildfire detection model at the base station to analyze pictures captured by LEO satellites and make predictions on the occurrence of wildfires. We use MobileNetV2 as our backbone model. MobileNetV2 is a lightweight model with low parameters. Lightweight models for image categorization offer faster inference, lower power consumption, efficient memory u sage, a nd ease of deployment, making them well-suited for real-time applications. However, a highly parameterized model can perform better than a poorly parameterized model. We apply transfer learning to ensure high performance in a low-parameterized model. The server must continuously check the images of all forest regions within a limited interval period to detect wildfires quickly. The lightweight model is very suitable for continuous image checking. Our experimental setup uses four baselines: VGG16, ResNet152, DenseNet161, and MobileNetv2. This study shows the efficacy of models in acquiring information from many sources within a given topic, as opposed to starting with no prior knowledge. We can also find from t he experimental results that our proposed low-parameterized transfer learning-based mobileNetV2 model achieves near 100 % accuracy, outperforming the other high-parameterized baselines.