Evaluation of burning detection using modified VGG19 for LULC classification changes

Abstract. This study aims to evaluate the effectiveness of a modified architecture of convolutional neural network (CNN) VGG19 for detecting fires and changes in land use and land cover classification (LULC). Remote sensing data from the Landsat 8 Operational Land Imagery (OLI) satellite was used to collect images from two distinct regions, one of which was used to obtain a dataset containing 1000 labeled images, and the other region was used to perform inference and verify the generalization of the model in an area with a high annual occurrence of fires. Analyses were conducted using a time series of normalized difference vegetation index (NDVI) and complementary cumulative distribution function (CCDF), to determine the potential for analysis in that area and define the periods of burning, pre-burning and post-burning. The VGG19 architecture was modified to maintain the input sizes of the images, resulting in a significant increase of 20.90 percentage points in the F1 score compared to the original architecture, as well as a 68.76% reduction in convergence time. In addition, the Gradient-weighted Class Activation Mapping (Grad-CAM) technique was used to improve the interpretability of the model at the moment of inference. The proposed methodology offers an approach for detecting burns by altering the LULC classification, and the modified VGG19 showed superior results.