The Cyprus island (9,251 k m 2 ) is located in the Mediterranean Sea and combines various ecosystems found across the basin. It has two prominent mountain ranges: the Troodos and the Kyrenia Mountains (also known as Pentadaktylos), with a central plain, the Mesaoria, lying between them. The peak of the range, Mount Olympus (also known as Mount Troodos), stands at an elevation of 1951 m. A significant portion of the island’s arable land, more than one-third, is irrigated and is primarily concentrated in the Mesaoria Plain and surrounding regions of Paphos in the southwest. This facilitates agricultural activities and contributes to the productivity of these lands.

As an input, the model receives RGB and NIR orthophotos collected over Cyprus during 2019 with a 10-cm resolution by the Department of Land and Survey of Cyprus. The GeoTIFF images used here are of high-resolution (10 cm per pixel), derived from UAV (drone) flight data with Red, Green, Blue (RGB), and Near-Infrared (NIR) bands, and use the ESRI:102,319 - CGRS-1993-LTM coordinate reference system (Transverse Mercator, meters). The images were processed using the Inpho GmbH software, commonly used in photogrammetry. The combination of RGB and NIR bands makes the images suitable for both visual interpretation and vegetation analysis. We split the dataset into small images, each of 0.25 k m 2 in size, that cover the Troodos, Athalassa and Aglantzia areas described above (428 images in total). The first step of the analysis was to produce specific training data by labeling trees in Cyprus. Since the model had already been pre-trained in Denmark, it was decided to use the method of transfer learning where the pre-trained model was fine-tuned using manually labeled tree crown data from Denmark and a new set of labels created for Cyprus (the size of the training set was determined by monitoring the performance of the model during the model training procedure aiming at a model loss of E−05 order - see Supplementary Appendix C). For the creation of the new Cyprus training dataset, trees from the each of the 3 studied areas have been labeled with 80% of labels in Troodos and 40% in Athalassa, and Aglantzia. As reference outputs, 6,000 individual tree crowns (mixed deciduous and coniferous) were delineated and labeled manually and further used together with 21,787 labels created for Denmark. This database was built through visual inspection of aerial images, without the aid of any semi-automatic assistance. The training areas were then created by considering rectangles, in which every tree has been delineated. The manual delineation of trees in Cyprus took approximately 2 months of computer-based processing and covered a variety of tree species and landscape types within urban and rural areas. This manual labeling was time-consuming but ensures that the reference database represents what the naked eye can see on an aerial image. No case separation was made for coniferous and broad-leaved trees. To distinguish neighboring tree crowns in dense forest areas, the spaces between adjacent crowns were included as input to the model, alongside the crown delineations. Note that this case of touching crowns is very rare in Cyprus due to the sparse distribution of trees, even in forests.

The deep-learning model of Li et al. (2023) to create maps of trees crown areas uses convolutional neural networks. Initially, it was developed to map trees in Denmark and further tested and validated in Finland and France (Li et al., 2023). Given the huge differences of trees and forest structure between Cyprus and Denmark, related to the dry climate of Cyprus which influences tree density, species and related crown shapes, as well as the presence of trees in mountains (not present in Denmark). The method used to re-tune the model for Cyprus is presented in the following section. We implemented tree counting and crown segmentation tasks simultaneously by using a multitask deep learning-based network (see Figure 2) derived from the U-Net architecture (Ronneberger et al., 2015), featuring two separate output branches dedicated to each task. The crown segmentation branch addressed a semantic segmentation challenge, where every pixel in an image was categorized either as part of an object (white) or as part of the background (black) (Wang, 2018). Therefore, the model outputs a binary mask (see Figure 3) showing tree areas with white pixels and non-tree areas with black pixels. The secondary branch estimates the overstory tree count by generating density maps through regression for a specific image. The ground truth density maps were created by applying normalized 2D Gaussian kernels (see Figure 2) on the delineated tree crowns (Zhang et al., 2016). The overall count of trees in an image of any dimension was determined by integrating the density map. Compared to counting by enumerating the segmented tree crowns, where several adjoining tree crowns might be incorrectly counted as one, the density estimation-based approach, according to the authors (Li et al., 2023), improves the overall counting bias.

The model architecture is a U-Net framework, widely employed in computer vision tasks. In line with the methodology proposed by Oktay (2018), the conventional U-Net was enhanced by integrating self-attention blocks to capture more relevant information during the down-sampling phase. Additionally, batch normalization was employed after each convolutional layer to enhance stability and accelerate the training procedure (Wang, 2018; Zhang et al., 2016). Most of the model parameters were shared between the two branches (segmentation and counting) with only those necessary for generating the ultimate output predictions being specific to each task. In the segmentation branch, a sigmoid activation function was employed in the final output layer to generate probabilities of a pixel to belong to a tree, ranging from 0 to 1. These probabilities were subsequently transformed into binary labels using a threshold of 0.5. According to Li et al. (2023), in the counting branch, a linear activation function is the best choice to preserve the values of the Gaussian kernel.

During each training epoch, random patches (tree crown labels) measuring 256 × 256 pixels were extracted from all accessible labeled images to create training and validation patches with a batch size of 8. These image patches were standardized, adjusting each instance and channel to have a mean of 0 and a standard deviation of 1, before being utilized as inputs to the network. The model was trained using the Adam optimizer for 2,500 epochs. Machine learning methods offer the opportunity to play with several parameters to improve the quality of the model. Several combinations of hyper-parameters (such as the optimizer, the number of epochs, etc.) were checked by inspecting the model performance, to find the most optimal one (see Supplementary Appendix A). The model was optimized by minimizing a loss function that was derived from the two branches and has been evaluated by monitoring closely the evolution of the loss function. The Tversky loss function (Salehi et al., 2017) was chosen as in Li et al. (2023). The model performed well for accurate estimate of tree counting (error of order 1 0 − 2 ). Various data augmentation methods were implemented, such as random flipping, cropping, Gaussian blurring, and brightness adjustment, to augment the dataset (Shorten and Khoshgoftaar, 2019).

For evaluation, we generated a new test dataset of 2,500 manually created tree crown labels, 2,000 labels in Troodos forest, and 500 labels of trees in Athalassa, and Aglantzia. There was no spatial overlap between this test data and the training data. There is unfortunately no accurate National Forest Inventory (NFI) sample data for Cyprus, as recent data based on permanent plots is lacking. Consequently, the evaluation of the model was done by using only our new testing dataset. For the crown area segmentation task, the model output was evaluated by calculating the common area between predictions and labels (F1 score). For the tree counting task, three metrics were used (Chalmers and Adkins, 2020; Sigal and Chalmers, 2016): the relative mean error (Equation 1), the relative total error (Equation 2), and the coefficient of determination R 2 (Equation 3), a measure of how well the independent variable explains the variability of the dependent variable. Relative MAE = 1 n ∑ i = 1 n y i − y ̂ i y ̄ (1) Relative total error RTE = ∑ i = 1 n y ̂ i − y i y i ∑ i = 1 n y i (2) R 2 = 1 − ∑ i = 1 n y i − y ̂ i 2 ∑ i = 1 n y i − y ̄ 2 (3) where y i , y ̂ i , and y ̄ denote the reference, prediction, and the mean reference value, respectively, and n denotes the total number of samples (tree crowns).

The relationship between tree cover density and the number of trees within an area allows us to better understand the structure of forests that include sparse trees. The correlation between these two variables for the Troodos National Forest Park, Aglantzia urban area, and Athalassa Forest Park is shown in Figures 6B, D, F, respectively.

For the Troodos Park, a correlation coefficient of 0.93 was found between tree cover density and the number of trees, as shown in Figure 6B. In other words, as tree cover density increases, there is a tendency for the number of trees to increase as well. The 1 k m 2 sub-areas with the largest TCD contains trees with the largest size of tree crown, equal to 23 m 2 . Over the Troodos Park, for a given number of trees (e.g., 3,000 per area) there is a large range of TCD. The area with minimum TCD (8%–10%) contains trees with an average crown size equal to 10 m 2 . This area has the potential to grow up to TCD of about 20% with an average tree crown size of 20 m 2 (as shown in Figure 6B). This information can be used in future studies to predict the potential growth of this area (e.g., in terms of carbon stock). The positive correlation suggests that the Troodos forest is quite homogeneous with a fairly constant crown area per tree. In specific areas where the above relationship does not apply, tree crown areas are different from the rest of the forest. This is consistent with the NFI data in Section 2.1.2 which indicates that the Troodos forest park encompasses a diverse array of tree species with different typical crown areas. Alternatively, the diversity of crown sizes within these areas can also be explained by a mix of large or mature trees next to smaller or younger ones. In Figure 6B, there are a few outliers that correspond to areas that accommodate a large number of trees with small tree crown sizes. We found that these outliers correspond to planted areas with many young trees at the Amiantos mines (see Supplementary Figure S8 in the Appendix section), within the Troodos forest park. This area has been planted in the context of a reforestation campaign of the Department of Forests between 2009 and 2014.

For the Athalassa National Forest Park, the situation is slightly different. The correlation between TCD and the number of trees across 1 k m 2 sub-areas is equal to 0.94 (see Figure 6D), possibly indicating either similar diversity of the tree species in this area with the Troodos forest or trees with similar maturity. This is consistent with the fact that the majority of the Athalassa forest park primarily consists of artificially planted trees for recreation and environmental conservation purposes.

For the urban area, which includes buildings, a large number of trees has been found. Despite the fact that a very small area has been considered as a sample, it was found that there is a strong positive correlation between TCD and tree numbers, equal to 0.99 (seen Figure 6F). This is a surprising result given the expected wide variety of trees (and related maturity) encountered within the city.

According to Figure 6D, the maximum tree cover density of a 1 k m 2 sub-area in Athalassa forest park is equal to 14 % , corresponding to almost 14,000 trees, with an average tree crown size equal to 10.5 m 2 . Such an area with tree cover density equal to 14 % can accommodate between 8,000 and 10,000 trees in the Troodos forest (Figure 6B), corresponding to an average tree crown size between 14 m 2 and 16 m 2 , respectively. In other words, for the same TCD, the Troodos forest has two times less trees but each with almost double crown size compared to Athalassa Park. Assuming these two forests are similar (in terms of abundance of tree species), such difference can indicate a different level of maturity for the two forests (with older trees in Troodos having larger crowns). Other features can also explain this difference such as the location of the forest park (in the mountains for Troodos and a flat terrain for Athalassa), and the limited amount of rainfall over the Athalassa area (compared to Troodos). For the urban area, it can be observed in Figure 6F that it contains a reference area of 1.0 k m 2 with a surprisingly high tree cover density, almost equal to 10 % , and many trees in it, almost 10,000 trees. A careful visual check of this reference area shows that it includes a green park with large trees, located in the urban area.

A well-known empirical law in closed forests, referred to as self-thinning, is a linear heuristic relationship between log-transformed tree density and individual tree crown size in a forest stand that reaches its maximum density (Farrior et al., 2016). Essentially, trees compete for resources and space, so that a given area can only accommodate either a few large trees or many small trees. This principle is well-established in dense forests (i.e., with high TCD) where trees compete for access to sunlight. It is therefore, very interesting to examine the law for non-closed (i.e., with low TCD) forests such as the Troodos forest park and the small Athalassa forest park, where trees do not compete for light but rather for water or nutrients. The self-thinning exponent for these two forests is calculated as the linear regression slope between the logarithm of the average area covered by a tree crown against the logarithm of the number of trees. For the Troodos Forest Park, the self-thinning exponent equals −0.45, indicating that as the average individual tree size increases, the number of trees decrease quickly, with a strong relationship between the two (see Supplementary Appendix E). This steeper slope suggests a higher level of competition among tree species for resources such as water and nutrients. It also implies a more pronounced hierarchical canopy structure, with larger, dominant trees more effectively competing for resources and a decrease in the number of smaller trees. Additionally, the maximum tree cover density in a reference area of Troodos was found to be approximately 45%, indicating that Troodos is a sparse forest. This observation is particularly interesting because the forest law considered here has not been previously applied to sparse forests. For the Athalassa forest park, the self-thinning exponent equals −1.17, which indicates a moderate to strong negative relationship between tree density and individual tree size within this small city forest park (see Supplementary Appendix E). This indicates that trees in the area compete strongly for resources and they form a relatively uniform canopy. Since the park consists of mainly man-made vegetation we could suggest that a lot (possibly too many) trees were planted per area in Athalassa compared to Troodos forest park (see slopes of Figures 6B, C. The self-thinning exponent carries significant implications for forest management practices. For example, it can help in making decisions related to thinning treatments, spacing of trees, and stand density management to optimize timber production, biodiversity conservation, or other management objectives.

Because it was not possible to get access to all the photos available for Cyprus, we established a transfer function between our very high resolution dataset and a coarser resolution dataset, available for the whole island. When considering different coarse resolution tree cover datasets in Cyprus, such as the Global Forest Watch (GFW), the European Forest Institute (EFI) Maps, the CORINE Land Cover (CLC), we decided to use the CORINE CLC + Backbone product with 10 m resolution for the year 2018 to derive a moderate resolution estimation of the total tree covered area over Cyprus (Copernicus, 2020). The CLC + Backbone raster product provides comprehensive land-cover mapping, distinguishing 11 different land cover types, one of which is the tree cover. It uses pixel-based, multi-temporal Sentinel-2 imagery to create a continuous 10 m resolution map. When overlapping land cover types occur within a single pixel (such as a combination of vegetation and sealed surfaces), the product assigns one dominant class, generally using a majority rule where the class covering more than 50% of the pixel is selected. In the case of tree cover, the pixels of tree cover class are classified into two forest type classes, coniferous and broad-leaved.

The reason behind the choice of CLC+ is the quality of the prediction, which for Cyprus is almost 90 % and its open access. The main differences between CLC + Backbone and our methodology and data sample are essentially i) the type of the sensors (CLC+: satellite images of Sentinel-2, our study: aerial images - orthophotos), ii) the resolution (CLC+: 10 m/pixel, Our study: 10 cm/pixel), iii) the size of the reference area (CLC+: all Cyprus, our study: 107 k m 2 ), and iv) the definition of tree cover (CLC+: categorizes each pixel to a land category and then to a forest type, our study: segments each tree without the space between them set as a condition the existence of the visible shadow of the tree to avoid including shrubs). The CLC + Backbone product forest type tree cover map was used in combination with our results to estimate the total tree covered area of Cyprus (occupied and non-occupied parts). Firstly, the tree covered area of our very-high-resolution dataset has been calculated over 428 images of 0.25   k m 2 . The lowest recorded tree cover density was 1.81% in an open, arid field with sparse tree coverage, while the highest tree cover density reached 44.94% in Troodos forest. As a first step, we resampled the CLC + data to match the resolution of our images ( 10 c m / p i x e l ) and re-projected it to the same coordinate system. After that, we cut the map into small parts with the same dimensions of our original data ( 0.25 k m 2 ) . As a next step, we calculated the percentage of each image that is forested (includes trees of any type). Now, for our studied areas we have the area covered by tree crowns per reference image area and we combine this information with the forested percentage. At this step we have a dataset of 106 images of 0.25 k m 2 for which we have the size of the area covered by tree crowns (output of our model) and the percentage of the image classified as forested in CLC+.

For the prediction of the size of the area covered by tree crowns per reference area, we used a Random Forest regression method, having as a feature value the forested percentage of CLC+ and as a target value the area covered by tree crowns from our very high resolution images. Using the area that is covered by our dataset of high-resolution images as a test set, we found that the performance of the model was good with Mean Absolute Error (MAE) equal to 0.0068, Mean Squared Error (MSE) equal to 9.125e-05 and R-squared (R ²): 0.82. The data in Figure 7) show that the model prediction is more efficient for areas with a moderate tree cover value compared to the rest of the data but also achieves the accurate prediction of tree cover in dense forest areas.

The RF results applied to the whole island indicate that in Cyprus, 1381 k m 2 are covered by tree crowns, which corresponds to almost 15 % of the total country area. According to Figure 8, when a very small part of a reference area is covered by tree crowns, the CLC + Backbone failed to identify it as being tree-covered. On the other hand, when a reference area includes a large fraction of tree crowns, the CLC + overestimated the tree cover. For instance, when we considered the Troodos forest park as a case study, we found that according to CLC+, 32.71 k m 2 are covered by trees compared only to 16.38 k m 2 in our very high resolution data. Thus, CLC + over-estimates tree cover by 100 per cent.