The problem of wildfire/ forest fire prediction is rather complex and has been approached from a variety of starting points. However, for models intended as tools for fire prevention organizations, the realism of rendering predictable dataset is important as it allows the results of prevention to be more readily interpretable. Also, for appropriate response to be affective it must be rendered in real time, the physically based models of fire remedies are computationally challenging.
Occurrences of wildfires prove that the WFAS (Wildland Fire Assessment System) is not connected with what is happening in the forests in real time. The SPC (Storm Prevention Center) is responsible for forecasting meteorological conditions which, when combined with the fuel conditions, favor rapid growth and spread of a fire, should an ignition occur. The dryness of the fuels is closely tied to both short and long term meteorological conditions. In the short-term, hot and dry atmospheric conditions can rapidly modulate the dryness of fuels that are available to burn. In the long-term, patterns that favor below normal rainfall and above normal temperatures can result in drought. Fuel dryness level forecasts are heavily relied upon for a baseline of where fuels are appropriately dry to warrant a highlighted area. Dead fuel moisture responds solely to ambient environmental conditions and is critical in determining fire potential. Dead fuel moistures are classified by time lag. A fuel’s time lag is proportional to its diameter and is loosely defined as the time it takes a fuel particle to reach two-third’s of its way to equilibrium with its local environment.
Most marketplace solutions are reactive; we aim for proactive solutions aimed at prevention rather than cure.
We have developed the Smart Wildfire Sensor device that can be used as a network of sensors to classify dead fuel images using Google’s TensorFlow in real time along with sensor data that would help predict the occurrence of a wildfire.
Until now, it was not possible to automatically recognize and classify various fuels (twigs, stems, branches, etc.) based on their sizes, although there are various manual mechanisms in place to get the fuel information (satellite, aerial or manual photography). With our Smart Wildfire Sensor, it is possible to classify the various fuels, in which, the computer learns how to classify them based upon initial learning phase that utilizes hundreds of sample images.
Training usually takes between 4-6 hours. Once trained, the Neural Network can classify a given fuel image with a confidence ratio. A 90% confidence indicates the proper bin in which the image has been placed. Once classified, SPC can quickly search for areas with dry fuel conditions in real time, along with other sensory data such as humidity, temperature, and pressure and wind speed, to estimate the likelihood of a wildfire.