Land use/Land cover mapping to monitor dead tree areas using multi-modal, multi-temporal Remote Sensing imagery and a deep learning model

Land Use/Land Cover (LULC) monitoring is essential for understanding Earth's surface dynamics, particularly in assessing the impact of vegetation and land changes on hydrological systems, vulnerability to extreme climatic events, and forest health. In recent years, increasing forest dieback caused by climate change, pests, and diseases has raised global concerns about ecosystem stability and biodiversity. The rapid spread of tree mortality and the need to accurately capture its temporal evolution highlight the necessity of precise detection and monitoring of dead tree areas, which are crucial for effective forest management and mitigating related environmental impacts.

To address this, our study aimed to map LULC and identify areas of dead trees from 2018 to 2023 in the Harz Mountains and its surrounding regions in Germany, an area severely affected by bark beetle infestation. For this purpose, we trained a multi-modal U-Net architecture, a supervised learning model with an encoder-decoder structure designed to capture contextual features across multiple scales. The training applied multi-temporal optical (Sentinel-2) and radar (Sentinel-1) imagery acquired during the growing season (May-August) of 2020-2021 as the training dataset, with ESA 2020/2021 data, tree species distributions from the Thünen Institute of Forest Ecosystems, and manually annotated dead trees as the reference dataset. Annual LULC maps for 2018-2023 were generated by processing each image using the trained model and subsequently combining the predictions per image using a majority voting approach, considering seven LULC classes: cropland, grassland, built-up areas, water bodies, coniferous, deciduous, and dead trees. Furthermore, a change analysis was performed on the predicted maps from 2018 to 2023.

Accuracy assessment demonstrated the model’s robust performance, with an overall accuracy of 0.88. Additionally, a comparison between a European LULC map -ELC10- and our predicted LULC map for 2018 resulted in an overall accuracy of 0.86, further highlighting the reliability of this method. Among the classes, cropland achieved the highest F1-score (0.97), likely due to the higher number of training samples available (40% of the total training samples). In contrast, the dead tree class demonstrated the lowest F1-score (0.60), attributed to its limited sample size (1% of the total training samples) and confusion with coniferous trees. The model effectively mapped the other classes, with F1-scores exceeding 0.70. The analysis revealed an increase in dead trees and grassland areas, primarily at the cost of coniferous trees, which can be linked to tree mortality caused by bark beetle infestation and prolonged drought, particularly from 2018 to 2022. It also revealed deforestation patterns between 2018 and 2023, with dead tree areas initially concentrated near Brocken in the Harz Mountains. Over time, these areas steadily expanded from the southeast towards the western and central parts of the study area.

These findings, based on freely accessible satellite data, can support forest managers in monitoring landscapes and tree mortality and help identify effective control measures.