Emma Turkulainen

Ms. Emma Turkulainen


Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
12 October 2023 in Remote Sensing
Verified
Journal Article
Open access
PDF file
ETEmma Turkulainen
Emma Turkulainen
EHEija Honkavaara
Eija Honkavaara
RNRoope Näsi
Roope Näsi
RORaquel Alves De Oliveira
Raquel Alves De Oliveira
THTeemu Hakala
Teemu Hakala
Päivi Lyytikäinen-SaarenmaaPäivi Lyytikäinen-Saarenmaa
Päivi Lyytikäinen-Saarenmaa
The widespread tree mortality caused by the European spruce bark beetle (Ips typographus L.) is a significant concern for Norway spruce-dominated (Picea abies H. Karst) forests in Europe and there is evidence of increases in the affected areas due to climate warming. Effective forest monitoring methods are urgently needed for providing timely data on tree health status for conducting forest management operations that aim to prepare and mitigate the damage caused by the beetle. Unoccupied aircraft systems (UASs) in combination with machine learning image analysis have emerged as a powerful tool for the fast-response monitoring of forest health. This research aims to assess the effectiveness of deep neural networks (DNNs) in identifying bark beetle infestations at the individual tree level from UAS images. The study compares the efficacy of RGB, multispectral (MS), and hyperspectral (HS) imaging, and evaluates various neural network structures for each image type. The findings reveal that MS and HS images perform better than RGB images. A 2D-3D-CNN model trained on HS images proves to be the best for detecting infested trees, with an F1-score of 0.759, while for dead and healthy trees, the F1-scores are 0.880 and 0.928, respectively. The study also demonstrates that the tested classifier networks outperform the state-of-the-art You Only Look Once (YOLO) classifier module, and that an effective analyzer can be implemented by integrating YOLO and the DNN classifier model. The current research provides a foundation for the further exploration of MS and HS imaging in detecting bark beetle disturbances in time, which can play a crucial role in forest management efforts to combat large-scale outbreaks. The study highlights the potential of remote sensing and machine learning in monitoring forest health and mitigating the impacts of biotic stresses. It also offers valuable insights into the effectiveness of DNNs in detecting bark beetle infestations using UAS-based remote sensing technology.

7 November 2024 in The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Verified
Journal Article
Open access
PDF file
ETEmma Turkulainen
Emma Turkulainen
JHJanne Hietala
Janne Hietala
JJJiri Jormakka
Jiri Jormakka
JTJohanna Tuviala
Johanna Tuviala
RORaquel Alves De Oliveira
Raquel Alves De Oliveira
EHEija Honkavaara
Eija Honkavaara
Climate change is threatening forest ecosystems worldwide by inducing various abiotic and biotic disturbances. In Europe, the European spruce bark beetle (Ips typographus L.) poses a significant threat, causing serious mortality in mature Norway spruce (Picea abies H. Karst.) stands. Rapidly evolving remote sensing technologies offer valuable tools for monitoring forest health, enabling timely management operations. This study presents a novel approach for large-area forest health monitoring using Uncrewed Aircraft Systems (UAS) and multispectral imaging. The research focuses on a hydrogen-powered Beyond Visual Line of Sight (BVLOS) airship for efficient monitoring of disturbances caused by I. typographus. A specific challenge is training machine learning models capable of covering wide areas. Our objective was to study the potential of deep learning models, including transfer learning and fine-tuning techniques, in developing the scalability and accuracy of UAS-based monitoring for detecting individual spruce trees and classifying their health. The approach was empirically evaluated in a study site in North Karelia, Finland. A multispectral image dataset was collected over a 1.3 km2 test area in May 2023 in a BVLOS setting operated from a command centre 75 km away. The results indicated that employing transfer learning significantly improved classification accuracy compared to training models from scratch, showing potential for implementing scalable machine learning methods for large-area UAS surveys. The best model yielded F1-scores of 0.936 for healthy, 0.955 for dead, and 0.817 for non-spruce classes. Furthermore, the results indicated that BVLOS airships offered high accuracy while reducing emissions and labour associated with UAS monitoring.

30 October 2025 in The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Verified
Journal Article
Open access
PDF file
ETEmma Turkulainen
Emma Turkulainen
RORaquel Alves De Oliveira
Raquel Alves De Oliveira
NKNiko Koivumäki
Niko Koivumäki
PLPäivi Lyytikäinen-Saarenmaa
Päivi Lyytikäinen-Saarenmaa
RNRoope Näsi
Roope Näsi
EHEija Honkavaara
Eija Honkavaara
Deep learning has significantly advanced forest health monitoring by enabling automated analysis of high-resolution aerial imagery. However, the generalization of these models across ecologically diverse regions remains limited due to domain shifts, which are variations in environmental conditions between training and testing locations. In this study, we propose a domain generalization (DG) framework that disentangles domain-invariant, task-relevant features from domain-specific environmental variations in multispectral UAS imagery. Our approach extends a baseline 2D convolutional neural network by incorporating parallel domain-specific and shared feature extractors, along with a domain classifier trained via adversarial learning. We evaluate the model using a leave-one-site-out strategy across three Finnish forest sites with diverse ecological characteristics. Results show that the DG model improves classification accuracy in previously unseen environments, with performance gains of up to 27% compared to the baseline. These findings highlight the effectiveness of feature disentanglement in enhancing the robustness and transferability of deep learning models for forest canopy health assessment, supporting more scalable and reliable forest monitoring solutions.

30 October 2025 in The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Verified
Journal Article
Open access
PDF file
AAAlva Anttonen
Alva Anttonen
APAxel Päivänsalo
Axel Päivänsalo
ETEmma Turkulainen
Emma Turkulainen
RORaquel Alves De Oliveira
Raquel Alves De Oliveira
KKKirsi Karila
Kirsi Karila
EHEija Honkavaara
Eija Honkavaara
Assessing the health of individual spruce trees in forests is critical for early detection of bark beetle infestations and effective forest management. This study presents a novel methodology that leverages multi-view uncrewed aerial vehicle (UAV) imagery to improve tree health classification at the individual tree level. High-resolution images were collected over a spruce-dominated forest and processed to extract multiple perspectives of each tree crown. Compared to the typical case of using one orthophoto per tree, our process yielded on average 31 images per tree crown. Deep learning models, including VGG16 and a simple CNN, were trained to classify trees as healthy, infested, dead, or non-spruce. Results demonstrate that incorporating multi-view images increased classification accuracy, particularly for the challenging infested and non-spruce categories, compared to traditional orthophoto-based approaches. The best-performing model achieved an overall accuracy of 0.94 and a macro F1-score of 0.85, with notable improvements in detecting infested trees.