Burned area detection based on Planet imagery using virtual SWIR band

Forest fires pose significant threats to both human safety and the natural ecosystem. Detecting and accurately estimating the extent of the burned area is crucial for effective response planning. Remote sensing emerges as a valuable solution for estimating the burned area, with various satellites employed in previous studies. Unlike these satellites, microsatellites offer a promising alternative with higher spatiotemporal resolution. In this study, we utilized PlanetScope imagery and implemented the U-Net model. PlanetScope provides images at a 3m spatial resolution and revisits the same area every day, offering a distinct advantage in accurately estimating burned areas across different scales of fire events. However, PlanetScope lacks a Shortwave Infrared (SWIR) band commonly used in forest fire studies. To address this limitation, a virtual SWIR band was introduced in this study. To enhance accuracy in specific regions, a virtual SWIR band was created using machine learning techniques using the SWIR images from Landsat and Sentinel-2. Our approaches were tested in four study regions. The U-Net model was employed to generate burned area prediction maps, and each model's performance was evaluated using several metrics, including intersection-over-union (IoU), mean IoU, recall, precision, F1-Score, and the Kappa coefficient. In this study, we not only validated the effectiveness of our proposed methods but also identified the potential to enhance the accuracy of burned area estimations, particularly for microsatellites lacking a SWIR band.