Urban green spaces provide essential ecosystem services that benefit community health. Analyzing tree coverage in urban areas at a detailed scale is crucial. Recently, deep learning models have been developed to estimate tree canopy height with high precision. This research evaluates the accuracy of two datasets in predicting tree coverage and canopy height in Milan. The aim is to assess their suitability for further use and improve methodologies for more precise urban green coverage analysis.

Introduction

Urban trees are a critical component of a city’s green infrastructure, providing many social, economic and environmental benefits (Lai & Kontokosta, 2019) in reducing greenhouse gas emissions, assisting with water run-off, sequestering carbon, and overall improving public health, thereby improving the quality of life for residents. Given the numerous external stressors facing cities today, urban areas are increasingly adopting nature-based solutions and advocating for a healthier urban forest as a means of building resilience for the future.

In this context, research has been conducted to obtain global datasets on tree coverage and height, using multiple deep learning models to estimate canopy structure from satellite images and aerial LiDAR datasets. This study aims to evaluate the suitability of two datasets in Milan: the ETH Global Sentinel-2 10m Canopy Height (2020) and the Global Canopy Height Maps, developed by Meta. It assesses the accuracy of green coverage through an analysis of NDVI values and compares the estimated canopy heights with the Tree Species Location dataset of the Comune di Milano.

Enabling Data and Methodology

Five datasets were used within this research, in particular:

• The Trees Species Location dataset, implemented and regularly updated by the Green Maintenance Unit of the Municipality of Milan in 2021, is a point based dataset, containing information on trees’ canopy height and width in Milan.
• Sentinel-2 satellite images with resolution 5x5m, captured on 19/06/2024, were used to calculate the normalized difference vegetation index (i.e., NDVI). This is an effective index for quantifying green vegetation. The value range of the NDVI is -1 to 1. Low, positive values represent shrub and grassland (e.g., approximately 0.2 to 0.4), while high values indicate temperate and tropical rainforests (e.g., values approaching 1) (Sentinel-2 L2A, n.d.).
• ETH Global Sentinel-2 10m Canopy Height (2020) is a global dataset with 10m ground sampling distance, based on a deep learning method for canopy height retrieval from Sentinel-2 optical images and GEDI LIDAR mission dataset (Lang et al., 2023).
• Global Canopy Height Maps, developed by Meta, is a global dataset with 1m ground sampling distance, based on satellite imagery spanning from 2009 to 2020, with a focus on data from 2018 to 2020 and built using AI models, as DiNOv2 (Nease, 2024).

Using these datasets, we developed a methodology to assess the accuracy of deep learning-based datasets—specifically, the ETH Global Sentinel-2 10m Canopy Height (2020) and the Global Canopy Height Maps—by comparing them against ground truth data, including Tree Species Location and Sentinel-2, NDVI index. We applied this methodology to Municipio 1 of the Municipality of Milan, where two of the city’s main parks are located.

To evaluate the accuracy of tree coverage, we overlaid the ETH Global Sentinel-2 10m Canopy Height (2020) and the Global Canopy Height Maps grids with the NDVI index values, creating a single vector layer based on the NDVI grid resolution. Finally, we calculated the confusion matrix for green coverage accuracy for the two deep learning datasets.

To assess canopy height accuracy, we refined the Tree Species Location dataset by applying a filter on the 5th to 95th percentile of its canopy height values, resulting in a height range of 1 to 30 meters. We then used the canopy radius value to estimate the extent of canopy coverage by buffering around the trees’ centroids and overlayed these values to the vector layer created in the previous step. This contained the NDVI values and the maximum canopy height values for the two deep learning datasets. Next, we filtered for areas with an NDVI value greater than 0.25, which represent the ground truth for green coverage. Finally, we compared the canopy height values derived from the Tree Species Location dataset to those computed for the two deep learning datasets.

Results

Results of this research outline the fitness for use of deep learning based tree canopy height datasets. Results are structured in two sections, outlining different methodologies for tree coverage and tree canopy height accuracy assessment.

Tree coverage analysis

Figure 2 and Figure 3 display the results of the tree coverage accuracy analysis. To perform this assessment, we overlaid the ETH Global Sentinel-2 10m Canopy Height (2020) dataset (see Figure 2) and the Global Canopy Height Maps (see Figure 3) grids with NDVI index values. We then computed the confusion matrix for both deep learning datasets. Specifically:

• True positives were calculated as the percentage of pixels where a deep learning dataset showed a height greater than 0, and the NDVI index was greater than 0.25;
• False positives were calculated as the percentage of pixels where a deep learning dataset showed a height greater than 0, and the NDVI index was less than or equal to 0.25;
• False negatives were calculated as the percentage of pixels where a deep learning dataset showed a height of 0, and the NDVI index was greater than 0.25;
• True negatives were calculated as the percentage of pixels where a deep learning dataset showed a height of 0, and the NDVI index was less than or equal to 0.25.

For the ETH Global Sentinel-2 10m Canopy Height (2020) dataset (see Figure 2), true positives account for 61% of the NDVI dataset, indicating that 61% of the areas where the NDVI value is greater than 0.25 are correctly identified. Conversely, false negatives constitute 39% of the areas where the NDVI value is greater than 0.25, corresponding to approximately 2km² of green areas that were not detected. The ETH Global Sentinel-2 10m Canopy Height (2020) dataset mistakenly identifies 6% of the areas in Municipio 1 without vegetation (i.e., NDVI ≤ 0.25) as green areas (i.e., False positives). This represents roughly 0.5 square kilometers of incorrectly detected green areas.

Figure 2 Confusion matrix spatial output, ETH Global Sentinel-2 10m Canopy Height (2020) dataset

For the Global Canopy Height Maps dataset (see Figure 3), true positives account for 59% of the dataset, thus 59% of the areas where the NDVI value is greater than 0.25 are correctly identified. In contrast, false negatives constitute 41% of the areas where the NDVI value is greater than 0.25, which corresponds to approximately 2.1km² of green areas that were not detected. Additionally, the Global Canopy Height Maps dataset incorrectly identifies 31% of the areas in Municipio 1 without vegetation (i.e., NDVI ≤ 0.25) as green areas (i.e., False positives), roughly 2.6km² of incorrectly detected green areas.

Figure 3 Confusion matrix spatial output, Global Canopy Height Maps dataset

Tree canopy height analysis

Figure 4 presents the results of the tree canopy height accuracy analysis. To conduct this analysis, we first cleaned the Tree Species Location dataset of the Comune di Milano and estimated the extent of canopy coverage by buffering around the trees’ centroids. We then overlaid the maximum canopy height derived from the dataset onto the vector layer containing only true positive values from the NDVI dataset and the two deep learning ones. Maximum values for canopy height were selected in the deep learning datasets, as these are computed using satellite images and therefore represent the top of the tree canopy.

Overall, there is a tendency for the ETH dataset to have a positive skew in height values compared to the Tree Species Location dataset. In contrast, the Global Canopy Height Maps mostly show negative differences when compared to the ground truth. For the ETH Global Sentinel-2 10m Canopy Height (2020) dataset, height differences range from -22 to +27 meters, while for the Global Canopy Height Maps dataset, height differences range from -29 to +30 meters. This indicates significant inaccuracies in both datasets, even in areas that overlap with the NDVI values.

Additionally, we found that 32% of the ETH dataset values show a height difference between -5 meters and +5 meters when compared to the Tree Species Location dataset, while the Global Canopy Height Maps dataset has 25% of its values within this range. This suggests that the ETH dataset has slightly better accuracy when estimating canopy height.

Lastly for the Global Canopy Height Maps dataset, we noticed an opposite trend in height values delta for parks and larger green areas, which mainly present negative height differences and smaller green areas, which conversely show positive height differences. This phenomenon is not presented in the ETH dataset.

Figure 4 Tree canopy height delta analysis results

Conclusions and Future Works

Urban trees are a critical component of a city’s green infrastructure, providing many social, economic, and environmental benefits (Lai & Kontokosta, 2019). Therefore, it is crucial to analyze tree coverage at a granular scale within urban environments. Recently, several deep learning models have been developed to estimate tree canopy height with high accuracy, achieving resolutions down to 1 meter. This research aimed to assess the accuracy of two deep learning datasets, ETH Global Sentinel-2 10m Canopy Height (2020) and Global Canopy Height Maps, in predicting tree coverage and determining tree canopy height in Municipio 1 of Milan.

To evaluate tree coverage accuracy, we used the NDVI index derived from Sentinel-2 satellite images to establish ground truth for the spatial location of green areas. Our findings show that both deep learning datasets perform similarly in correctly detecting existing green areas (i.e., True Positives), but the Global Canopy Height Maps dataset frequently predicts green areas where none exist (i.e., False Positives). For assessing canopy height accuracy, we compared the datasets against the Tree Species Location dataset from the Comune di Milano, which served as ground truth. However, we observed significant difference in the estimated heights for both datasets.

The comparison methodology has several limitations, including reprojection errors and the challenges associated with using the Tree Species Location dataset, which is point-based and required buffering to estimate canopy width. Additionally, comparing these datasets with satellite images could introduce inaccuracies. Future work will focus on refining the assessment methodology to leverage open-source data and models for more precise analysis of urban green coverage and canopy height.

---

Acknowledgments

We thank the Municipality of Milan for sharing data. The analyzed data were treated according to the GDPR-General Data Protection Regulation (EU, 2016/679). This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

References

Lai, Y., & Kontokosta, C. E. (2019). The impact of urban street tree species on air quality and respiratory illness: A spatial analysis of large-scale, high-resolution urban data. Health & Place, 56, 80–87. https://doi.org/10.1016/j.healthplace.2019.01.016

Lang, N., Jetz, W., Schindler, K., & Wegner, J. D. (2023). A high-resolution canopy height model of the Earth. Nature Ecology & Evolution, 7(11), 1778–1789. https://doi.org/10.1038/s41559-023-02206-6

Nease, J. T., Camille Couprie, John Brandt, Justine Spore, Tobias Tiecke, Tracy Johns, Patrick. (2024, April 22). Using Artificial Intelligence to Map the Earth’s Forests. Meta Sustainability. https://sustainability.atmeta.com/blog/2024/04/22/using-artificial-intelligence-to-map-the-earths-forests/

Sentinel-2 L2A. (n.d.). Retrieved September 2, 2024, from https://docs.sentinel-hub.com/api/latest/data/sentinel-2-l2a/