Performance Analysis of Deep Convolutional Autoencoders with Different Patch Sizes for Change Detection from Burnt Areas

Abstract

Fire is one of the primary sources of damages to natural environments globally. Estimates show that approximately 4 million km2 of land burns yearly. Studies have shown that such estimates often underestimate the real extent of burnt land, which highlights the need to find better, state-of-the-art methods to detect and classify these areas. This study aimed to analyze the use of deep convolutional Autoencoders in the classification of burnt areas, considering different sample patch sizes. A simple Autoencoder and the U-Net and ResUnet architectures were evaluated. We collected Landsat 8 OLI+ data from three scenes in four consecutive dates to detect the changes specifically in the form of burnt land. The data were sampled according to four different sampling strategies to evaluate possible performance changes related to sampling window sizes. The training stage used two scenes, while the validation stage used the remaining scene. The ground truth change mask was created using the Normalized Burn Ratio (NBR) spectral index through a thresholding approach. The classifications were evaluated according to the F1 index, Kappa index, and mean Intersection over Union (mIoU) value. Results have shown that the U-Net and ResUnet architectures offered the best classifications with average F1, Kappa, and mIoU values of approximately 0.96, representing excellent classification results. We have also verified that a sampling window size of 256 by 256 pixels offered the best results.

Uni-Temporal Multispectral Imagery for Burned Area Mapping with Deep Learning

“…Bermudez et al [55] used a conditional Generative Adversarial Network to synthesize missing remote sensed optical data from Sentinel-1 SAR data for a region with the presence of burned area. Recently, de Bem et al [56] analyzed the performance of deep convolutional autoencoders (U-Net and ResUnet) using bitemporal image pair of the Landsat scenes and recommended the sampling window size of 256 by 256 pixels in DL model training.…”

“…Moreover, the EMSR447, EMSR298_05, and EMSR298_03 products provide the reference data for the test sites in Corinthia, Fågelsjö-Lillåsen, and Trängslet, respectively. Differently, for the Elephant Hill and Enskogen fires, their dNBR images, calculated from cloud-free pre-fire and post-fire Sentinel-2 images, were empirically thresholded to elaborate the precise ground truth mask within the official perimeters from the Copernicus EMS (EMSR298_01) and BC Wildfire Service (K20637) [67] as de Bem et al [56] did. Furthermore, we manually refined all the burned area annotations based on visual analysis of VHR post-event optical images (i.e., Google Earth Map).…”

Instance Segmentation for Large, Multi-Channel Remote Sensing Imagery Using Mask-RCNN and a Mosaicking Approach

“…DL enables pattern recognition in different data abstraction levels, varying from low-level information (corners and edges), up to high-level information (full objects) [4]. This approach achieves state-of-the-art results in different applications in remote sensing digital image processing [5]: pan-sharpening [6][7][8][9]; image registration [10][11][12][13], change detection [14][15][16][17], object detection [18][19][20][21], semantic segmentation [22][23][24][25], and time series analysis [26][27][28][29]. The classification algorithms applied in remote sensing imagery uses spatial, spectral, and temporal information to extract characteristics from the targets, where a wide variety of targets show significant results: clouds [30][31][32][33], dust-related air pollutant [34][35][36][37] land-cover/land-use [38][39][40][41], urban features [42][43][44][45], and ocean [46][47]…”

“…This is a data limitation for optical Earth observation sensors that are generally multispectral, where the available channels provide complementary information that maximizes accuracy. In semantic segmentation, approaches to aggregate more information considered: (a) the use of image fusion techniques, where the three bands used are data integration products [84]; (b) input layer adequation to support a larger amount of channels, e.g., 14 channels [15] 12 channels [14], 7 channels [85], and 4 channels [86].…”

“…Albuquerque et al (2020) evaluate the segmentation's accuracy, considering sliding windows with different overlapping strides. Research has also been carried out to evaluate different sliding window sizes [14,18,89]. The frames' fixed size must consider a dimension that allows the general context to perform the classification without a significant increase in computational complexity and CNN parameters.…”