Extending deep learning approaches for forest disturbance segmentation on very high‐resolution satellite images. Issue 3 (26th January 2021)

• Record Type: Journal Article
• Authors: Kislov, Dmitry E.; Korznikov, Kirill A.; Altman, Jan; Vozmishcheva, Anna S.; Krestov, Pavel V.
• Editors: Disney, Mat; Cord, Anna
• Is Part Of: Remote sensing in ecology and conservation. Volume 7:Issue 3(2021)
• Page Start: 355
• Page End: 368
• Publication Date: 2021-01-26
• Subjects: Bark beetle outbreak -- DCNN -- deep convolutional neural network -- deep learning -- forest damage detection -- vegetation recognition; Remote sensing -- Periodicals; Ecology -- Research -- Periodicals; Ecology -- Methodology -- Periodicals; Ecology -- Remote sensing -- Periodicals; Nature conservation -- Methodology -- Periodicals; 577.0723
• DOI: 10.1002/rse2.194
• Languages: English
• ISSNs: 2056-3485

Abstract

Abstract: Accurate remote detection of various forest disturbances is a challenge in global environmental monitoring. Addressing this issue is crucial for forest health assessment, planning salvage logging operations, modeling stand dynamics, and estimating forest carbon stocks and uptake. Substantial progress on this problem has been achieved owing to the rapid development of remote sensing devices that provide very high‐resolution images. Concurrently, image processing algorithms have witnessed rapid development owing to the extensive use of artificial neural networks with complex architectures and deep learning approaches. This opens new opportunities and perspectives for applying deep learning methods to solving various problems in environmental sciences. In this study, we used deep convolutional neural networks (DCNNs) to recognize forest damage induced by windthrows and bark beetles. We used satellite imagery of very high resolution in visual spectra represented as pansharpened images (RGB channels). When predicting forest damage, we obtained accuracies higher than 90% on test data for recognition of both windthrow areas and damaged trees impacted by bark beetles. A comparative analysis indicated that the DCNN‐based approach outperforms traditional pixel‐based classification methods (AdaBoost, random forest, support vector machine, quadratic discrimination) by at least several percentage points. DCNNs can learn a specific pattern of the area of interest and thus yield fewer false positive decisions than pixel‐based algorithms. The ability of DCNNs to generalize makes them a good tool for delineating smooth and ill‐defined boundaries of damaged forest areas, such as windthrow patches. Abstract : We demonstrated that the proposed deep learning algorithm (U‐Net‐like CNN) is an efficient method of automatically recognizing forest sites disturbed by winds and bark beetles. If appropriately trained, a U‐Net‐like CNN can identify specific types of damaged forests and their locations. In contrast with standard machine learning methods, deep learning algorithms do not require complex feature engineering. They are able to discover pattern‐specific features internally and yield good recognition results with images of different resolutions and captured in slightly different conditions. The main advantage of deep learning is the ability to understand the surrounding context of areas of interest. The use of deep learning algorithms in such cases reduces the number of incorrectly segmented pixels owing to the ability to learn the surrounding context.