Automatic Windthrow Detection Using Very-High-Resolution Satellite Imagery and Deep Learning

Abstract

Wind disturbances are significant phenomena in forest spatial structure and succession dynamics. They cause changes in biodiversity, impact on forest ecosystems at different spatial scales, and have a strong influence on economics and human beings. The reliable recognition and mapping of windthrow areas are of high importance from the perspective of forest management and nature conservation. Recent research in artificial intelligence and computer vision has demonstrated the incredible potential of neural networks in addressing image classification problems. The most efficient algorithms are based on artificial neural networks of nested and complex architecture (e.g., convolutional neural networks (CNNs)), which are usually referred to by a common term—deep learning. Deep learning provides powerful algorithms for the precise segmentation of remote sensing data. We developed an algorithm based on a U-Net-like CNN, which was trained to recognize windthrow areas in Kunashir Island, Russia. We used satellite imagery of very-high spatial resolution (0.5 m/pixel) as source data. We performed a grid search among 216 parameter combinations defining different U-Net-like architectures. The best parameter combination allowed us to achieve an overall accuracy for recognition of windthrow sites of up to 94% for forested landscapes by coniferous and mixed coniferous forests. We found that the false-positive decisions of our algorithm correspond to either seashore logs, which may look similar to fallen tree trunks, or leafless forest stands. While the former can be rectified by applying a forest mask, the latter requires the usage of additional information, which is not always provided by satellite imagery.

Extending deep learning approaches for forest disturbance segmentation on very high‐resolution satellite images

“…In the case of the windthrow detection problem, it is sufficient to use images that cover several tens of meters, as this is usually enough to recognize fallen stems on images. For pansharpened satellite imagery with a resolution of 0.3-0.6. meter/pixel, it is usually sufficient to take 256 9 256 images (Kislov & Korznikov, 2020).…”

“…We did not use random shifts in the augmentation pipeline. Such The Pretrained U-Net CNN was developed using snapshots from Airbus Pleiades-1A/B images with spatial resolution 0.5 m/pixel; for details, see supplementary materials in Kislov and Korznikov (2020). transformations would be redundant because sub-images were cropped out from a fixed set of satellite images and often intersected each other that could be considered as they are spatially shifted.…”

Estimating VAIA Windstorm Damaged Forest Area in Italy Using Time Series Sentinel-2 Imagery and Continuous Change Detection Algorithms

“…Remote sensing (RS) is considered the most efficient way to map forest disturbances over large area. Many RS data were tested to map and estimate forest windstorm disturbances as for example Airborne Laser Scanner data (ALS) [8], Sentinel-1 SAR data [13,18], high and medium spatial resolution multispectral satellite imagery [15,[18][19][20]. However, some of these approaches present relevant limitations.…”

“…The one using SAR data proposed by Rüetschi et al (2019) [13], is not able to determine the exact boundary of damages that is mandatory to estimate the damaged area and cannot be applied for winter storm and/or in absence of a high-resolution digital terrain model (i.e., 2 m resolution). The approaches proposed by Kislov & Korznikov (2020) [19] and Hamdi et al (2019) [20] using high resolution imagery and deep learning seems to be promising in mapping the boundaries of windstorm disturbances; however, the cost of high resolution images over large areas is high.…”

Automated Bird Counting with Deep Learning for Regional Bird Distribution Mapping

“…The methods applied in the discussed papers vary from traditional image analysis techniques including thresholding and template matching to segmentation and supervised/unsupervised classification, but still, no deep learning method is mentioned. Even though papers related to machine learning applications on unmanned aerial vehicle- (UAV) or satellite-derived images have been published so far [ 43 , 44 , 45 , 46 , 47 ], the studies focusing on on-ground photography are limited [ 48 , 49 ]. At the same time, digital photography has entered a new era with the availability of cloud-storage and location-aware cameras and smart-phones that enable the sharing of photographs on a truly massive scale.…”