Classification of Tree Species in Poland Using CNNs Tabular-to-Pseudo Image Approach Based on Sentinel-2 Annual Seasonality Data

Abstract

Tree species classification provides invaluable information across various sectors, from forest management to conservation. This task is most commonly performed using remote sensing; however, this method is prone to classification errors, which modern computational approaches aim to minimize. Convolutional neural networks (CNNs) used to model tabular data have recently gained popularity as a highly efficient classification tool. In the present study, a variation of this method is used to classify satellite multispectral data from the Sentinel-2 mission to distinguish between 18 common Polish tree species. The novel model is trained and tested on data from species-homogeneous forest stands. The data form a multi-seasonal time series and cover five years of observations. The model achieved an overall accuracy of 80% and Cohen Kappa of 0.80 of the raw output and increased to 93% with post-processing procedures. Considering the large number of species classified, this is a promising and encouraging result. The presented results indicate the importance of early vegetation season reflectance data in model training. The spectral bands representing the infrared, red-edge and green wavelengths had the greatest impact on the model.

1. Introduction

Remote-sensing-based tree species classification provides valuable information for a variety of sectors ranging from forest management to conservation [1,2,3,4]. The steadily growing availability of airborne and satellite sensor data coupled with the increasing processing power and availability of cloud-based computational services [5,6] greatly facilitates the classification process. Recently, the process is most often entrusted to machine learning algorithms of different architectures. The choice of the classification algorithm is often determined by the character of the dataset—whether it is multispectral or point cloud, tabular or images. Among the popular approaches to tackle the classification problem, random forest (RF) [7,8,9,10] and support vector machine (SVM) [11,12,13] are the most popular when the data is in tabular format. Imaging data are often processed using a variety of convolutional neural networks (CNNs) adaptations [14,15,16].

With respect to Earth observation datasets, species classification tasks are now often based on Sentinel-2 mission imagery due to its suitable resolution, frequent revisits and relative ease of access [8,17,18,19,20,21,22]. Due to spatial and time resolution limitations, the use of imagery from the Landsat mission is gradually becoming less popular [11,19,23,24], but new applications are still being proposed [25] with a large spatial scale analyses in mind. There are also successful studies using high-resolution WorldView-2 [7,16,26], WorldView-3 [12] and Formosat-2 [27] satellite imagery. A separate group of studies are classifiers based on hyperspectral data from airborne vehicles (drones, aircraft) [13,28] or commercially available datasets [29].

While standard machine learning algorithms, such as random forest and SVM, have been widely utilized in forest classification due to their robustness and speed [10,11], recently, deep learning approaches, particularly CNNs, have emerged as powerful tools, demonstrating superior performance in various remote sensing tasks, including tree species classification [30,31,32,33,34]. This enhanced capability is largely attributed to their capacity for automated feature learning, enabling CNNs to iteratively learn relevant data transformations and hierarchical features directly from raw input, thereby making traditional, heuristic feature engineering obsolete. Furthermore, CNNs are uniquely adept at exploiting spatial patterns through successive convolutions and pooling operations, providing a crucial advantage for analyzing complex remote sensing imagery and significantly enhancing their effectiveness, especially with very high spatial resolution data [35]. Recent innovative CNN (convolutional neural network) -based methods have progressed beyond discrete classification to quantify tree species proportions at the pixel level, allowing for mapping across all forest compositions from pure to highly mixed stands. This approach provides a more realistic representation of heterogeneous forest environments, optimizing the use of available forest inventory data that may not offer pixel-level labels [30].

The approach proposed in this study—transforming tabular data into pseudo-images—offers a promising pathway for leveraging the powerful capabilities of (CNN), which are inherently designed for spatial data and typically cannot be directly applied to tabular structures due to the lack of inherent locality and spatial relationships. In the case of analyzing medium-resolution satellite imagery such as Sentinel-2, the standard approach is to treat them as tabular data rather than as image data in the sense used by CNN approaches. To the best of our knowledge, our study represents the first example of transferring an approach—converting tabular data into two-dimensional CNN—that has already been tested in other fields [36,37] to the context of remote sensing.

In the current paper, a classifier capable of discriminating 18 common tree species (

Table 1. List of tree species selected to analyze. For each species, the table lists the abbreviation, name, number of 10 × 10 m cells, and number of forest stands with Sentinel-2 satellite information.

Abbreviation	Latin Name	Popular Name	Cells Sampled	Number of Stands
ASH	Fraxinus excelsior L.	European ash	196,470	336
BEE	Fagus sylvatica L.	European beech	458,448	315
BLC	Robinia pseudoacacia L.	Black locust	167,099	351
BPN	Pinus nigra Arn.	Black pine	65,240	161
DMP	Pinus mugo Turra	Dwarf mountain pine	117,172	127
EBA	Alnus glutinosa (L.) Gaertn.	European black alder	506,234	520
FIR	Pseudotsuga menziesii (Mirb.) Franco	Douglas fir	71,295	178
HBM	Carpinus betulus L.	European Hornbeam	86,110	176
LAR	Larix decidua Mill.	European larch	331,760	300
OAK	Quercus undefined	Oak undefined	398,095	360
PIN	Pinus sylvestris L.	Scots pine	436,300	355
ROA	Quercus rubra L.	northern red oak	162,934	278
SBR	Betula pendula Roth	Silver birch	332,170	355
SFR	Abies alba Mill.	Silver Fir	506,935	303
SLL	Tilia cordata Mill.	Small-leaved linden	57,385	127
SPR	Picea abies (L.) H.Karst	European spruce	231,753	269
SYC	Acer pseudoplatanus L.	Sycamore maple	50,063	163
WPP	Populus alba L.	White poplar	292,989	584

) based on spectral data from the Sentinel-2 mission [38] is presented. The classifier was trained and evaluated on data from homogeneous (species-wise) forest stands scattered over the entire territory of Poland. The data were pixel-based in tabular format, but the choice of the modelling algorithm was influenced by the recently popular formula of using CNN on tabular data transformed into pseudo-images [39,40]. The architecture of the proposed two-dimensional CNN is inspired by one-dimensional CNN, which achieved high accuracy in Kaggle competition in classification of tabular data [41]. Additional inspiration was an approach using a one-dimensional CNN on time series data proposed by Xi [42].

This paper presents a modern approach to tree species classification in the hope of promoting state-of-the-art modelling algorithms offering better classification accuracy than commonly used ensemble learning models. The presented study introduces a comprehensive classification framework that significantly advances the scope of tree species identification by enabling the accurate classification of a broad range of species. This contrasts with prior approaches, which have predominantly concentrated on limited, species-specific datasets [8,19,42]. By addressing a wider taxonomic spectrum, this methodology represents a substantial improvement in automated classification, with implications for enhanced scalability and applicability across diverse ecological and forest management contexts.

2. Materials and Methods

2.1. Data and Preprocessing

The study area spanned an extensive region across Poland, encompassing a vast area from 14°07″ to 24°09″ east longitude and from 49°00″ to 54°50″ north latitude. The data was obtained from the area belonging to the State Forests, which were covered with forest. The reference data was derived from the Forest Data Bank (FDB) database. The FDB online service (https://www.bdl.lasy.gov.pl/portal accessed on 15 June 2025) provides free access and the ability to download the data set with a multitude of forest stand variables, including species composition, age, average height of species, average diameter at breast height (DBH), and so forth. The data are created during preparation of forest management plans and are annually updated. The data extracted for this study included information on the geometry of the forest stands, the species composition, the age of the species, and the site type. A total of eighteen tree species were selected for analysis. The list of species and the number of stands chosen for analysis is presented in Table 1. Only single-species stands with an area exceeding 0.4 hectares were selected for analysis. The age of the stands ranged from 5 to 160 years. For species where the number of available stands in the database was less than 500, all stands were used for creating a reference dataset. For species where the number of available stands was between 500 and 1000, 500 stands were randomly selected. For species with more than 1000 available stands in the database, 600 polygons per species were randomly drawn using stratification. The stands were grouped by age class (eight classes of 20-year intervals) and site type (21 classes). Then, for each group within a species, the same number of stands were drawn in the stratification group to give a final count of 600 polygons for that species. Initially selected polygons for all species were subsequently verified manually on the basis of an up-to-date orthophoto map with a resolution of 0.25 m, which was available on the national geoportal (geoportal.gov.pl). During the on-screen verification, the boundaries of the stands’ polygons were edited if another species were identified, for example, close to a stand boundary. If more than one species was identified within the polygon boundaries and it was not possible to clearly define the range of a particular species, such polygons were removed and not considered as reference.

After manual verification a total of 5258 reference stands were selected for further analysis. The number of chosen reference stands by species is summarized in Table 1. The location of the reference forest stands used in the analysis is shown in Figure 1. Based on the data set, training and test sets were prepared.

The data from the Sentinel-2 satellite for the period from 2018 to 2022 were acquired through Google Earth Engine [5]. For each year, four mosaics were generated for a two-month interval covering the months of April to May, June to July, August to September and October to November, which we refer to as “seasons”. Cloud masking was conducted via the s2cloudless algorithm (Sentinel-Hub) [43]. Mosaics were generated by computing the median value of cells from overlapping Sentinel-2 scenes. Mosaics for each season were prepared from 10 spectral bands, including B2-B8a and B11-B12. A raster layer containing 40 bands (4 seasons * 10 spectral bands) was created for each year.

To ensure uniformity across seasons, the layers were standardized using the calculated mean and standard deviation, according to the following formula: z = (x − μ)/σ [44].

Next, the values of raster cells (observations) were extracted from the prepared spatial layers for the 5258 reference fields. To reduce the influence of outliers, observations below the 5th percentile and greater than the 95th percentile were excluded from the data set. Similarly, spatial layer cells marked as missing data were also excluded from further analysis.

This sampling procedure resulted in 4,468,452 valid pixel observations (896,596 for the year 2018, 896,651 for the year 2019, 882,366 for the year 2020, 896,183 for the year 2021 and 896,656 for the year 2022). Table 1 shows the number of observations for each tree species.

Next, the input dataset was randomly divided into training and testing sets in a 75%/25% ratio. When dividing the input data set, care was taken to ensure that all data from a reference polygon, but from different years, were stored either within the training set or the test set, and were never mixed.

Each observation in the datasets represents a time series with location coordinate information appended. The time series covers the reflectance variation during the growing season in one year. It provides insight into the annual seasonality. The five years studied allow a broader variability to be captured.

Figure 2 shows the mean values of standardized reflectance by tree species and season. Each line represents the average reflectance magnitude in the different Sentinel-2 bands used in analysis.

2.2. Method

The data were organized in a tabular format. A convolutional neural network (CNN) was employed to analyze them. CNNs are primarily utilized for image analysis. The novel approach of applying such networks to tabular data within the context of image classification is used in this study. The workflow of the conducted analysis is presented in Figure 3a and is described in the following sections of the paper.

2.2.1. CNN Topology

The CNN architecture employed in this research is inspired by one-dimensional CNNs that have achieved high levels of accuracy in classifying tabular data [41]. However, a two-dimensional CNN architecture demonstrated superior performance on this particular dataset and was ultimately utilized in the study. The model was implemented in the Tensorflow [45]/Keras [46] environment. The topology of the network used is illustrated in Figure 3b.

The 40-dimensional input data vector was fed into a 100-dimensional fully connected layer, which employed a rectified linear unit activation function to expand the input. The layer was subsequently transformed into a 2D pseudo-image representation comprising 10 × 10 pixels. A set of two convolutional layers were subsequently implemented; the first comprised 32 filters (kernel size 2 × 2, stride 1, and 0 padding, a Rectified Linear Unit (RELU) activation function), while the second incorporated 64 filters (kernel size 3 × 3, stride 1, and 0 padding, RELU activation function). The output was then flattened and subjected to regularization through the introduction of a 0.5 dropout. A fully connected layer (32 nodes, RELU activation function) and a fully connected output layer (18 nodes, softmax activation function) were subsequently added. The number of layers and associated hyperparameters were initially tuned using the KerasTuner optimization framework [47] and further adjusted experimentally based on model performance. The Adam optimization algorithm was employed to learn the model. In an attempt to mitigate model overfitting, a relatively modest learning rate (0.00001) and an early-stopping mechanism were implemented. The learning curves of the CNN designed to classify tree species from tabular data are presented in Figure 4.

The effect of input variables on the model output was examined using SHAP index [48]. The magnitude of importance is a measure of how strongly a variable influences the model outcome [49].

2.2.2. Post-Processing

Two approaches were tested to improve tree species recognition. In the first approach, a species was selected based on network responses above 0.8. In the second approach, results were combined due to belonging to a single reference field and a score was assigned when all network responses indicated the same species. The first approach favored quality over quantity. It allowed for a significant increase in accuracy. However, at the same time, this approach reduced the number of classified pixels. The second approach is limited to single-species forest stands.

2.2.3. Model Quality Assessment

The following measures were used to assess model quality: Overall Classification Accuracy (CA), Cohen Kappa Index (CK), Precision, Recall, F1 Score, Receiver Operating Characteristics curve (ROC), Area under the ROC curve (AUC) and Confusion Matrix.

3. Results

The final CNN model achieved high performance with overall classification accuracy of 0.803 and Cohen’s Kappa of 0.804. All performance metrics are presented in

Table 2. Model performance metrics calculated for species.

Species	Accuracy	Precision	Recall	F1 Score	Support
ASH	0.56	0.54	0.56	0.55	37,356
BEE	0.86	0.76	0.87	0.81	80,108
BLC	0.85	0.80	0.86	0.83	32,165
BPN	0.55	0.77	0.56	0.65	18,165
DMP	0.96	0.96	0.97	0.96	35,749
EBA	0.85	0.87	0.86	0.86	169,852
FIR	0.55	0.74	0.56	0.64	17,040
HBM	0.55	0.67	0.55	0.61	19,620
LAR	0.88	0.84	0.89	0.86	57,875
OAK	0.73	0.77	0.73	0.75	114,673
PIN	0.92	0.85	0.93	0.89	112,260
ROA	0.69	0.69	0.69	0.69	35,600
SBR	0.81	0.80	0.82	0.81	82,715
SFR	0.91	0.86	0.91	0.88	121,970
SLL	0.40	0.48	0.40	0.44	7830
SPR	0.75	0.83	0.75	0.79	57,179
SYC	0.34	0.46	0.35	0.39	9918
WPP	0.69	0.77	0.69	0.73	84,900

. The ROC curves (with AUC values) and the confusion matrix are shown in Figure 5 and Figure 6, respectively.

The values of the resulting SHAP index for each input variable are shown in Figure 7. The figure provides only the 20 variables that have the most significant impact on changing the CNN’s response. The variables are ordered in descending order of strength of influence. The most influential variables were bands representing shortwave infrared, green and vegetation red edge wavelengths for the early vegetation season (April–May).

In the current study, modeling was particularly successful in recognizing four conifer species: dwarf mountain pine, Scots pine, silver fir, and European larch. Similarly well classified were selected deciduous species: robinia pseudoacacia, European beech, European black alder and silver birch.

However, there were tree species that were more difficult to classify (see Figure 5 and Figure 6 and Table 2 for details). The poorest performance was observed for sycamore maple, for which the classifier confused with silver birch and European black alder. The second worst species was small-leaved linden, which was most often confused with European black alder, oak, European ash, and silver birch. This was followed by European hornbeam, which was misclassified with a variety of deciduous species. The other problematic species were Douglas fir. This species was misclassified as either silver fir or Scots pine. Black pine was mistaken for Scots pine.

Despite these classification issues, the overall accuracy and Cohen’s Kappa are over 0.8. That shows the potential of the approach based on tabular data analysis.

The classifier successfully distinguished between deciduous and coniferous trees in 96% of cases. It is worth noting that there is no significant relationship between the sample size of the test set (see Table 2 column: Support) and the classification score of the species (Spearman’s Rho = 0.22); thus, the performance of the classifier is most likely due to the characteristics of the spectral response of the species encoded internally by CNN as an artificial image.

The map of classified tree species for the whole country is presented in Figure 8.

Post-processing procedures have made it possible to improve the accuracy of the model output. The overall accuracy was increased in two proposed scenarios. The results are as follows: after selecting the output above the maximum probability threshold of 0.8, the accuracy reached 0.932; after polygon grouping, the accuracy of the classifier reached 0.835.

4. Discussion

The tree species used in the study cover the most common species in Poland. There was no particular key for species selection, and the 18 species selected were the best represented in the available dataset (FDB), which ensured proper model training.

Seasonal changes in canopy spectral response are a strong discriminator of tree species [50,51,52,53] and were included in the current study as a sequential set of repeated measurements of spectral variables averaged over four two-month intervals from April to November. The length of the two-month intervals was determined by the quality and acquisition frequency of the satellite imagery. In this way, the composites provided nearly cloud-free mosaic coverage of the study area. The observations were a seasonal time series. The spectral response is also expected to vary between years due to various meteorological and biological factors [54,55]. Capturing this variability by including satellite images from five consecutive years was intended to introduce a climate influence into the model.

During data pre-processing, the training and test datasets were prepared to eliminate spatial autocorrelation. Data from a given training polygon were assigned exclusively to either the training or test dataset, regardless of the date of acquisition. In addition, the test dataset was normalized based on the statistics of the training dataset, both to promote the model’s ability to classify upcoming, unseen data.

The training polygons were deliberately chosen to represent the stands with homogeneous tree species. Such an approach allows for an efficient model training process. However, there were strong outliers in the raw spectral data set. Two approaches were tested to resolve them. The first approach excluded polygons with a variance of more than 2 σ in any of the variables in a given year, while the second approach used Winsorization and showed a positive effect on model performance. Therefore, the second procedure was selected as the outlier treatment for the study.

CNNs are typically designed to extract information from images where spatially meaningful patterns are expected [56]. They are most commonly used for computer vision and object classification tasks. Recently, CNNs have been shown to outperform regular deep learning models when trained on tabular data preprocessed into a pseudo-image format [39,40]. The choice of CNN dimensionality in the case of tabular data is determined less by the original data format and more by performance metrics. In the current study, a two-dimensional CNN model was used after preliminary testing. The basic architecture of the model was based on the well-performing competition architecture [41]. The model architecture was adapted to the input dataset and roughly hyperparameterized using the automatic hyperparameter optimization tool [47]. The implementation of the MaxPooling layer was also tested due to its reported positive effect on classification results [57]. It was ultimately removed due to a lack of positive impact on model performance. The model used in this study appears complex compared to the relatively small dataset and could be prone to overfitting, but thanks to strong regularization with a dropout layer, a low learning rate, and an early stopping mechanism, it was possible to benefit from the model’s good feature extraction ability while maintaining its ability to generalize.

Complicated machine learning models are difficult to interpret [58]. The use of the SHAP values method [48,59] revealed factors influencing the presented model. The variables that most discriminated species were those representing the shortwave infrared (bands B11 and B12), red edge (bands B5 and B8a), and green (band B3) wavelength spectra (Figure 7). This result is expected and reassuring that the models are working correctly. This is because the red edge is postulated to be specifically sensitive to the chlorophyll content of vegetation (see [60] for references). Short-wave infrared has been reported to be sensitive to leaf water content [61], vegetation density [62], and nitrogen content [63]. In addition, the “green” spectral region has been reported to aid in the detection of chlorophyll content where “red” bands fail [64]. The early growing season (April–May) clearly stands out as the most influential on the model (Figure 2 and Figure 7), most likely due to the influence it can have on the separation between deciduous and coniferous species [8].

In order to extract meaningful information from the model output, certain post-processing steps are usually required. In this research, two basic solutions are presented. Both solutions depend on different practical scenarios of model application. For the classification of mixed-species or unmanaged forests, the most useful approach seems to be to set the threshold of the output maximum probability. A supervised classifier assumes that the class with the highest probability value is assigned as the most likely output. However, if the data does not match the classes learned by the model, the highest probability value may still be relatively low. This indicates the inability of the model to classify with confidence. In such a case, setting a probability threshold (e.g., 0.8) and considering only outputs above this threshold helps ensure that the assigned class is not a near-random guess. However, this approach reduces the number of output datasets. In the case of spatial analysis, this will result in leaving empty (No Data) spatial cells. However, this approach increased the accuracy from 0.804 (raw output) to 0.932 (post-processing probability threshold) in the study presented. Another potential post-processing solution used in this study was to calculate the mode of the outputs associated with a coherent single-species area. This approach seems to be useful only in scenarios where monocultures are studied, and has the advantage of eliminating data noise. This approach, where the output mode is calculated for test polygons regardless of the year of data collection, improved the classification results, increasing the accuracy from 0.804 (raw output) to 0.835 (area mode post-processing).

By converting tabular data into image formats, various approaches aim to incorporate two-dimensional spatial information that CNNs can process efficiently, potentially leading to improved model performance [36,37,65]. A significant advantage observed in several studies is the enhanced classification accuracy when using the transformation to tabular data to pseudo-image. For instance, the Low Mixed-Image Generator for Tabular Data method, which transforms tabular data into 2D grayscale images, combined with CNNs, demonstrated superior area under the ROC curve values for cardiovascular disease prediction compared to traditional machine learning (ML) models like least absolute shrinkage and selection operator (LASSO) and Tabular Prior Fitted Network (TabPFN), especially when transfer learning was applied [36]. Similarly, the Novel Algorithm for Convolving Tabular Data (NCTD) approach, which uses mathematical transformations to create synthetic images, consistently surpassed conventional ML algorithms in classification accuracy across seven out of ten diverse datasets [65]. Furthermore, the Feature Clustering-Visualization (FC-Viz) method, by structuring tabular data as color-encoded images with preserved feature relationships, achieved the highest average test accuracy compared to SVM, RF, Decision Tree, and eXtreme Gradient Boosting across ten benchmark datasets [66].

While some studies reported that CNNs on pseudo-images could outperform machine learning algorithms, it is also noted that the improvements might not always be statistically significant or consistently superior to the best-optimized machine learning models [37]. This suggests that while conversion adds spatial information, it might not always sufficiently equip the CNN to grasp underlying patterns with significantly greater accuracy than highly optimized typical machine learning models based on tabular models.

Beyond accuracy, the pseudo-image approach often offers significant advantages in parameter efficiency and model size, leading to faster training and inference times [66]. However, it is also acknowledged that some pseudo-image methods might generate large image representations, which could increase computational demands for storage and processing, potentially limiting scalability in resource-constrained environments [65].

The general trend from studies conducted in other fields indicates that a well-designed tabular-to-image conversion can yield models that are not only competitive in predictive performance but also more efficient in terms of computational resources and model footprint. In our study, we transferred experience from the analysis of medical images to remote sensing applications. Although this was not the primary aim of our study, future research should focus on comparing the proposed approach with standard machine learning methods based on tabular data, which are commonly used for tree species classification based on medium-resolution satellite images.

5. Conclusions

This study presents a modern approach to tree species classification using a CNN architecture designed for tabular data—a method that is gaining popularity in the field. The proposed CNN model demonstrated strong performance in classifying 18 tree species in Poland, utilizing publicly available satellite imagery data from the Sentinel-2 mission. To enhance the model’s accuracy, specific post-processing steps were recommended, resulting in an accuracy of 0.93 in this study. Regarding data collection, images captured during the early vegetation season proved to be the most valuable for this classification task. Additionally, infrared, red-edge, and green wavelengths were identified as the most influential in enabling the model to effectively distinguish between tree species.