Assessment of vegetation indices for mapping burned areas using a deep learning method and a comprehensive forest fire dataset from Landsat collection

Forested areas, crucial for their multifunctional roles, face significant risks from wildfires. Recent methods have enabled accurate delineation of burned areas (BA) using satellite images, which is essential for assessing impacts and recovery. Nevertheless, many such techniques rely on short time series for training data and utilize many predictors without thorough analysis of efficient computational frameworks. Therefore, this paper presents a novel method that builds a forest fire dataset while minimizing technical and technological costs, while also assessing the benefits of incorporating vegetation spectral indices (VIs) versus specific spectral bands in a Convolutional Neural Network (CNN) detector. The methodology involves creating a dataset of real forest fires from 1985 to 2021 using two consecutive Landsat images and applying VIs and unsupervised clustering techniques. The dataset is then used to explore the feasibility of integrating VIs from single-temporal Landsat images into a U-Net-based CNN, measuring performance across multiple settings. A real dataset was successfully acquired, achieving a 73.68% level of agreement with the limited years available in the Spanish Ministry’s official records, while the analysis of the introduction of VIs into a CNN revealed that using single-temporal images with the Landsat blue band and VIs achieved a 73% Intersection over Union (IoU) metric, leading to a 46% reduction in storage requirements. In contrast, using the green band with the blue band and VIs resulted in a 75% IoU and a 31% reduction in storage, comparable to configurations using visible and infrared bands without VIs. Consequently, this methodology successfully collects a dataset of BA from wildfires, extending the digitized time scale while demonstrating the advantages of applying VIs, rather than specific spectral bands, in a CNN for delineating burned areas, achieving data optimization and normalization.