Automated Linear Disturbance Mapping via Semantic Segmentation of Sentinel-2 Imagery

TL;DR

A deep convolutional neural network model based on the VGGNet16 architecture is employed for semantic segmentation of lower resolution Sentinel-2 satellite imagery, creating precise multi-class linear disturbance maps in Alberta, Canada, demonstrating the effectiveness of the VGGNet model for accurate linear disturbance retrieval.

Abstract

In Canada's northern regions, linear disturbances such as roads, seismic exploration lines, and pipelines pose a significant threat to the boreal woodland caribou population (Rangifer tarandus). To address the critical need for management of these disturbances, there is a strong emphasis on developing mapping approaches that accurately identify forest habitat fragmentation. The traditional approach is manually generating maps, which is time-consuming and lacks the capability for frequent updates. Instead, applying deep learning methods to multispectral satellite imagery offers a cost-effective solution for automated and regularly updated map production. Deep learning models have shown promise in extracting paved roads in urban environments when paired with high-resolution (<0.5m) imagery, but their effectiveness for general linear feature extraction in forested areas from lower resolution imagery remains underexplored. This research employs a deep convolutional neural network model based on the VGGNet16 architecture for semantic segmentation of lower resolution (10m) Sentinel-2 satellite imagery, creating precise multi-class linear disturbance maps. The model is trained using ground-truth label maps sourced from the freely available Alberta Institute of Biodiversity Monitoring Human Footprint dataset, specifically targeting the Boreal and Taiga Plains ecozones in Alberta, Canada. Despite challenges in segmenting lower resolution imagery, particularly for thin linear disturbances like seismic exploration lines that can exhibit a width of 1-3 pixels in Sentinel-2 imagery, our results demonstrate the effectiveness of the VGGNet model for accurate linear disturbance retrieval. By leveraging the freely available Sentinel-2 imagery, this work advances cost-effective automated mapping techniques for identifying and monitoring linear disturbance fragmentation.

Figures (6)

• Figure 1: Study area in Alberta, Canada (denoted in red), encompassing the Boreal Plains (below the blue boundary) and Taiga Plains ecozones (above the blue boundary).
• Figure 2: Visualization of Sentinel 2 imagery and ABMI linear disturbance labels. The color scheme is as follows: blue for roads, yellow for pipelines, and red for cutlines. a) Original Sentinel-2 imagery. b) Sentinel-2 imagery overlaid with ground-truth ABMI vector labels. c) Sentinel-2 imagery with ABMI labels rasterized based on features touching the pixel center. d) Sentinel-2 imagery with ABMI labels rasterized when features touch any part of the pixels.
• Figure 3: The VGGNet-16 architecture takes a $4\times224\times224$ image as input and outputs a $4\times224\times224$ per-class prediction map, from which a final $1\times224\times224$ label map is obtained using the argmax operation. Blue blocks denote $3\times3$ convolutional layers, with labels indicating the number of channels. Green lines denote a batch normalization layer followed by a ReLU activation function, red lines denote $2\times2$ max pooling layers, yellow lines denote bilinear up-sampling by a factor of 2, $\oplus$ denotes summation operation, and the magenta line denotes a hyperbolic tangent activation function.
• Figure 4: Visualization of Sentinel-2 imagery, ground-truth labels, and predicted labels for three test tiles. Each row represents a different test tile, which was not seen during training. The first column shows the original Sentinel-2 imagery. The second column displays the corresponding ground-truth labels, and the third column shows the predicted labels generated by the model. In the ground-truth and predicted label images, yellow denotes pipelines, blue denotes roads, and red denotes cutlines.
• Figure 5: Confusion matrices normalized: a) by rows to show recall along the diagonal and b) by columns to show precision along the diagonal.