Overcoming Structural and Environmental Challenges in Identifying Brazilian Savanna Tree Species Using Deep Learning

Overcoming Structural and Environmental Challenges in Identifying Brazilian Savanna Tree Species Using Deep Learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Overcoming Structural and Environmental Challenges in Identifying Brazilian Savanna Tree Species Using Deep Learning Jéssica Silva Reis, Marcos Henrique André Deus, Carlos Silva-Neto, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8841518/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The Brazilian Cerrado, a global biodiversity hotspot, faces persistent challenges in species identification due to the limitations of phenology‑dependent methods in its seasonally dynamic landscapes. This study evaluates the performance of convolutional neural networks (CNNs) for bark‑image‑based classification of three ecologically prominent tree species: Apuleia leiocarpa , Astronium fraxinifolium , and Vochysia haenkeana . We compiled 1,515 bark images from individual trees (DBH ≥ 25 cm) during the 2023–2024 rainy seasons and applied data augmentation and normalization. Using the MobileNetV2 architecture, we trained and validated the model with metrics including Accuracy, Precision, Recall, F1‑score, Confusion Matrix, ROC/AUC curves, and t‑SNE projections. The model achieved an overall accuracy of 90.52%. Bark morphological complexity strongly influenced classification: V. haenkeana , with distinct patterns, showed the highest performance (Precision 1.00, Recall 0.94), while A. fraxinifolium and A. leiocarpa , which share more convergent bark traits, exhibited higher misclassification rates (22.54%). These results demonstrate how interspecific bark variability affects CNN discrimination and confirm that intrinsic bark heterogeneity (e.g., rhytidome texture, rugosity, color patterns, scars) and environmental variation increase classification difficulty. Our findings highlight the potential of bark‑based deep learning models as phenology‑independent tools for large‑scale forest inventories and biodiversity monitoring in complex ecosystems. A key limitation is the dataset’s restriction to a single seasonal period, underscoring the need for broader temporal sampling. This study reinforces the role of deep learning in delivering scalable and accurate solutions for ecological research and conservation in understudied biodiversity hotspots. Morphological Texture Analysis Environmental Monitoring Image Classification Models Tropical Ecosystems Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction South American savannas represent a significant portion of global savanna coverage, spanning over 2.29 million square kilometers and distinguishing themselves as the wettest and most biodiverse spectrum globally (Lira-Martins et al. 2022 ; Rocha and Pinto 2021 ; Schwaida et al. 2023 ). The Brazilian Cerrado, a globally recognized biodiversity hotspot, harbors an extraordinary diversity of endemic species essential for ecosystem functioning and resilience (Lira-Martins et al. 2022 ; Rabeling et al. 2019 ). Its structural and compositional complexity creates a mosaic of ecological niches, supporting exceptionally high levels of functional diversity (Lira-Martins et al. 2022 ; Rocha and Pinto 2021 ; Schwaida et al. 2023 ). Yet, despite its ecological prominence, the Cerrado remains critically understudied, posing substantial challenges for accurate species identification and evidence-based conservation planning (Colli et al. 2020 ; Mustin et al. 2017 ; Silveira et al. 2025 ). Indeed, the comprehensive understanding of biodiversity is frequently hampered by “knowledge shortfalls”, where crucial research questions remain unanswered due to data deficiencies (Hortal et al. 2015 ). Such deficiencies are often exacerbated by the fact that traditional approaches, such as classical taxonomy and botanical field surveys, though foundational, face inherent challenges in providing comprehensive and efficient large-scale identification (Arora et al. 2025 ; Rzanny et al. 2019 ). These challenges arise especially due to the psychophysiological complexity of seasonal forest formations and logistical constraints that hinder professional access to remote areas (Meyer et al. 2016 ). In response to these substantial identification challenges that hinder effective biodiversity management, novel technological solutions are rapidly transforming ecological assessment and monitoring. Among these, convolutional neural networks (CNN’s) have emerged as remarkably powerful and tools for large-scale species monitoring (Li et al. 2020 ; Zhang et al. 2024 ). Their unparalleled ability to analyze complex spatial, temporal, and spectral data with high precision stems from their design as deep learning models, capable of automatically identifying intricate hierarchical patterns - including edges, textures, and shapes - thereby circumventing the need for extensive manual feature engineering (Aggarwal 2018 ; Warner et al. 2024a ). This inherent proficiency has driven their widespread adoption in diverse ecological applications, ranging from accurate species identification and habitat mapping to critical environmental monitoring, such as deforestation detection and assessment(Kim et al. 2022 ). Furthermore, CNN-based models have demonstrated significant potential for forest inventories, enabling rapid and accurate classification of plant species from remote sensing data and successfully integrated into drones and scanning systems for detailed insights into inaccessible forested areas (Nooralishahi et al. 2021 ; Wu et al. 2021 ). Despite the well-established efficacy of CNN’s in various Classification tasks, a significant research gap persists regarding their specific application for native Cerrado tree species. Identifying these species through bark morphology offers a non – destructive approach, especially valuable when key phenological traits are seasonally inconspicuous (Kim et al. 2022 ). This is crucial, as botanical surveys frequently encounter abundant sterile vegetative material, and species identification in the Cerrado is further complicated by intraspecific variation across diverse habitats (e.g., forests, open fields, various soil types) (Blaanco et al. 2016 ; Fekri-Ershad 2020 ). Furthermore, bark offers high consistency throughout seasons and easy accessibility, even in high crown conditions (Rosell et al. 2014 ). However, its full potential for automated identification remains critically underexplored for species within semideciduous seasonal forest formations (Fekri-Ershad 2020 ). This gap is critical given the unique complexities of this biome and the urgent need for enhanced conservation tools. While expert botanists can often distinguish these species based on a combination of traits, complementary technological tools, including AI-driven systems, are increasingly essential to support field identification efforts (Caccianiga et al. 2021 ; Enríquez-de-Salamanca). Such tools can provide rapid, large-scale assessment, crucial for enhancing existing botanical knowledge and aiding conservation amidst rapid savanna destruction and climate change. This aligns with the broader development of interactive identification keys and comprehensive digital databases. This inherent difficulty in identification stems from subtle morphological variations, similar bark coloration, and shared intrinsic bark features (e.g., rhytidome texture, rugosity, color patterns, and scars) among co-occurring species (Gama et al. 2025 ; Surendran et al. 2025 ). This challenge is especially exacerbated during dry seasons and fires, when key phenological traits are less available for accurate diagnosis(Lira-Martins et al. 2022 ; Rabeling et al. 2019 ). Such challenges make specialized knowledge and precise identification tools limiting students in early stages of training and non-specialized individuals. For instance, environmental plasticity leads to confusion even for seemingly distinct species: Vochysia haenkeana , despite its marked yellow coloration, is often mistaken for Albizia niopoides by non-specialists. Similarly, Apuleia leiocarpa and Astronium fraxinifolium present highly similar rhytidome patterns and variable coloration in different environments (e.g., more brownish in closed forest vs. typical ashen/dark yellow in open environments), leading to frequent misidentification. Furthermore, these three species – V. haenkeana , A. leiocarpa , and A. fraxinifolium - can be confused with each other by the local community, as their bark tends to become more yellowish and subtly smooth depending on the time of year and climatic conditions, intensifying the ambiguity in visual identification. This ambiguity directly impacts the sustainable use of the Cerrado rich flora. Many species, particularly those in semideciduous seasonal forest formations, possess high economic potential (e.g., valuable timber, edible fruits, medicinal properties, non-timber forest products) (Berte et al. 2024 ; Silva-Pereira et al. 2011 ). However, their sustainable exploitation and valorization are severely hindered by this ongoing identification ambiguity, preventing adequate economic valuation and limiting community engagement in sustainable forest management (Souza et al. 2016a ). Consequently, this contributes to the undervaluation and displacement of native resources by expanding agricultural frontiers, further exacerbating biodiversity loss (Rouhan and Gaudeul 2014 ). Given the proven social acceptance of CNN’s among farmers and increasing initiatives for savanna plants thorough these networks are highly promising (Hajjaji et al. 2025 ; Nogueira et al. 2019 ). This allows producers to identify valuable species without confusion, fostering sustainable economic development and empowering local communities within the Cerrado. To directly address this pressing research and practical challenge, this study investigates the applicability of advanced deep learning techniques. Specifically, we evaluate the MobileNetV2 convolutional neural network for classification of three ecologically prominent Cerrado tree species - Apuleia leiocarpa (Fabaceae), Astronium fraxinifolium (Anacardiaceae), and Vochysia haenkeana (Vochysiaceae) - based solely on bark images. These species were selected for their ecological significance, widespread occurrence, and the morphological complexity of their bark surfaces, which present a nuanced and realistic challenge for automated recognition systems, as highlighted by the field identification difficulties and their economic/ecological value. Our research is guided by two central questions, which also represent key methodological challenges in this context: Firstly, we aim to understand how the inherent morphological complexity of bark influences CNN classification accuracy, particularly when differentiating between species exhibiting simpler versus more intricate patterns. Secondly, we investigate how the model’s performance is influenced by natural variability of intrinsic bark characteristics. This includes assessing the impact of features integral to itself – such as rhytidome texture and rugosity, color patterns, and the presence of scars – which can act as confounding variables for the classifier. By systematically exploring these questions, this study aims to promote CNN-based species identification methodologies and advance the application of deep learning in biodiversity conservation within conservation within complex ecosystems. Our findings will not only underscore the promising role of artificial intelligence in ecological research but also transparently delineate the practical limitations and highlight crucial future directions for developing highly accurate and robust CNN-based identification systems tailored for challenging ecological contexts. Methods Field Study and Data Collection This study employed a dataset specifically acquired to investigate bark texture patterns of tree species in the Brazilian savanna (Cerrado). A total of 1,515 images were systematically collected, evenly distributed among three native tree species within this biome: Apuleia leiocarpa (Vogel) J.F. Macbr. (Fabaceae), Astronium fraxinifolium Schott (Anacardiaceae), and Vochysia haenkeana Mart. (Vochysiaceae). Crucially, each species contributed 505 unique images, with each image representing a distinct individual tree, all with a Diameter at Breast Height (DBH) equal to or greater than 25 cm. Images were captured during the rainy seasons of 2023 and 2024, in mesophytic forests located in southern Goiás State, Brazil, specifically the municipalities of Piracanjuba (17°18'30.42"S, 49°1'44.37"W) and Abadia de Goiás (16°45'25.44"S, 49°25'35.12"W). Representative examples of the bark morphology for each species are presented in Fig. 1 , including both general trunks. A map illustrating the spatial distribution of the sampled individuals is provided in Fig. 2 . The selection of these species was previously justified in the Introduction based on their ecological significance, widespread occurrence, and the morphological complexities of their bark surfaces, which present a nuanced challenge for automated identification systems. This selection strategy was integral to our experimental design, allowing us to evaluate the model’s performance across a spectrum of bark morphological distinctiveness, from more uniform to subtly convergent patterns. For image acquisition, two types of equipment were utilized to ensure diversity and quality in the records: a Nikon D90 camera (equipped with an 18–105 mm lens, 4288 x 2848 pixels resolution), and the integrated camera of a Samsung Galaxy A71 smartphone (4000 x 3000 pixels resolution). All images were stored in JPEG format. Acknowledging the inherent variability of natural light incidence and the dynamic influence of canopy openness in understory environments, efforts were made to standardize image capture. During collection, the imaging device was consistently positioned perpendicular to the tree trunk, at distance ranging from 20 cm to 40 cm, depending on the conditions found in the field (Gama et al. 2025 ). Image capture was performed under natural lighting conditions, specifically between 9:00 AM and 3:00 PM, to mitigate extreme variations in light intensity and shadow effects. The residual variations in lighting and other environmental factors inherent to field data were specifically addressed during the dataset preprocessing through robust data augmentation techniques (as detailed in Section Dataset Preparation and Preprocessing), designed to enhance the model’s resilience to diverse real-world conditions. Machine Learning Pipeline Overview The overall methodology for bark image classification, encompassing data preparation, model training, and evaluation, is schematically presented in Fig. 3 . This pipeline was designed to rigorously test our research questions regarding the influence of bark morphology and superficial elements on CNN classification accuracy. The following subsections detail each component of this pipeline. Dataset Preparation and Preprocessing Upon collection, all images were initially in JPEG format and organized using the Hierarchical Dataset Organization (Xu and Goodacre 2018 ). To optimize feature extraction and computational efficiency for CNN training, and to enhance the detection of local texture patterns, the original high-resolution bark images were segmented into multiple smaller sub-images, or patches (Altal et al. 2025 ). From each original image, 60 random patches of 256 x 256 pixels were extracted, a number considered adequate according to previous studies utilizing between 9 and 80 patches (Elgamily et al. 2025 ; Misra et al. 2020). This resulted in a total dataset of 90,900 patches for analysis. This patch-based approach is particularly beneficial for training convolutional neural networks as it increases dataset variability and focuses the model on fine-grained textural details. This patch-based approach is particularly advantageous for training convolutional neural networks, as it increases dataset variability and optimizes the model's focus on fine-grained textural details (Baihaqi et al. 2025 ). The dataset was systematically divided into stratified training, validation, and test sets to ensure robust model development and unbiased evaluation (Liu et al. 2025 ). For each species, 85% of the images were allocated for model training and validation, while the remaining 15% were exclusively reserved for the test set (Cicero et al. 2017 ; Shah et al. 2020 ). This partitioning ensures that the model's generalization performance is evaluated on unseen data, mitigating potential overfitting( Altal et al. 2025 ) . To further enhance the model's ability to generalize across diverse real-world lighting and orientation conditions, and to effectively prevent overfitting, data augmentation techniques were extensively applied to the training set (Shorten et al. 2021 ). These techniques included randomized adjustments such as rotation (± 20°), zoom (± 20%), horizontal flipping, and brightness modifications (± 15%). Prior to model training, all images underwent a crucial normalization process (Kumar et al. 2024 ), where pixel values were rescaled to the [0, 1] range. This preprocessing step ensured data uniformity, mitigated potential illumination variances, and guaranteed compatibility with the MobileNetV2 model’s input requirements (Indraswari et al. 2022 ). Machine Learning The MobileNetV2 architecture, a highly efficient convolutional neural network, was selected for its proven capability in detecting relevant features in images, making it suitable for deployment on mobile and edge devices (Gulzar 2023 ). This architecture, configured with transfer learning techniques, leverages a model pre-trained on the extensive ImageNet database(Akay et al. 2021 ). This pre-training provides an initial set of robust recognition features(Al-Gaashani et al. 2025 ), such as edges, textures, and shapes, which are then fine-tuned to the specific characteristics of our bark image dataset. Hyperparameter optimization was a critical step to enhance the model’s robustness and reliability (Kumaresan et al. 2021 ). Beyond standard parameters like learning rate, number of epochs, and batch size, the optimization process aimed to fine-tune the model's capacity to automatically extract and differentiate salient morphological patterns from bark images. This approach enabled the model to effectively adapt to and discern subtle morphological specificities and potential visual noise inherent in our dataset, directly addressing aspects related to our research questions concerning bark complexity and superficial elements(Altal et al. 2025 ). Qualitative characteristics of bark morphology, including external coloration, texture, thickness, shedding pattern, color change with shedding, and the presence of superficial elements (e.g., lichens, bryophytes, scars), were visually assessed during dataset preparation (Fig. 4 ). This visual assessment provided crucial insights for experimental design and aided in diagnosing model performance and classification challenges. Model Evaluation and Statistical Analysis Prior to model evaluation, and to investigate potential spatial autocorrelation that could affect subsequent analyses, a Moran's I analysis was conducted for key intrinsic bark characteristics, including rhytidome texture, bark dimension (length, width, thickness, and weight), rugosity, color patterns, and scars. This analysis, performed using the Geoda software v.1.8.10 (Anselin et al. 2005 ), revealed no significant spatial autocorrelation for these variables, suggesting their spatial independence (Moran et al. 2021 ). The MobileNetV2 classification model's performance was rigorously evaluated on an independent test set using standard metrics, providing insights crucial for addressing our research questions. These metrics included Accuracy (ACC), representing overall correct predictions; Precision (P), measuring the quality of positive predictions; Recall (R), quantifying the model's sensitivity to identify all positive instances; and the F1-score (F1), the harmonic mean of Precision and Recall, indicating a balanced performance. Their calculations are defined as: $$\:ACC=\frac{TP+TN}{TP+TN+FP+FN}$$ $$\:P=\frac{TP}{TP+TN}$$ $$\:R=\frac{TP}{TP+FN}\:$$ $$\:F1=2*\frac{P*R}{P+R}$$ Where TP, TN, FP, and FN represent True Positives, True Negatives, False Positives, and False Negatives, respectively(Leiva-Bianchi et al. 2025 ; Yang et al. 2025a ). The confusion matrix was employed to systematically visualize classification performance across species. This matrix was critical for assessing the model's ability to distinguish species with similar bark characteristics(Fan 2025 ; Miftahushudur et al. 2025 ), particularly A. fraxinifolium and A. leiocarpa , thereby directly informing our investigation into how morphological complexity influences CNN classification accuracy. It also facilitated the identification of misclassification patterns potentially linked to visual noise from superficial elements, contributing to our understanding of the impact of such confounding features on model performance(Miftahushudur et al. 2025 ). For further interpretation of the model’s discriminative capabilities, Receiver Operating Characteristic (ROC) curves and their Area Under the Curve (AUC) were generated for each species (Huo and Glickman 2025 ). Additionally, t-Distributed Stochastic Neighbor Embedding (t-SNE) was applied to visualize high-dimensional feature vectors in two dimensions, revealing data clusters and patterns of similarity or overlap among species in the latent space (Mohammad et al. 2025 ). These visualizations provided qualitative insights into how the model differentiates, or struggles to differentiate, species based on bark morphology and the impact of visual noise, supporting the interpretation of our research findings. Plant Material The three native Cerrado tree species - Apuleia leiocarpa (Fabaceae), Astronium fraxinifolium (Anacardiaceae), and Vochysia haenkeana (Vochysiaceae) - were selected due to their ecological significance, widespread occurrence, and high economic potential in semideciduous seasonal forest formations (Souza et al. 2016b ). Their identification presents significant challenges stemming from subtle morphological variations, variable bark coloration (influenced by environment and season), and shared superficial features (e.g., coloration, texture, and scars). This complexity leads to frequent misidentification by non-specialists, with other co-occurring species like Albizia niopoides and among themselves ( A. leiocarpa , A. fraxinifolium , and V. haenkeana ) (Martello et al. 2023 ). These inherent ambiguities severely hinder the sustainable exploitation and valorization of these valuable resources, ultimately contributing to the undervaluation and displacement of native resources by agricultural expansion(Martello et al. 2023 ). Consequently, these species serve as ideal test cases for evaluating AI-driven tools in addressing real-world identification bottlenecks in complex and economically vital ecosystems. Results The MobileNetV2 model demonstrated strong performance in classifying the three Cerrado tree species from bark images. The model’s performance metrics and visual analytics, including the confusion matrix, Receiver Operating Characteristic (ROC) curves, a heatmap of bark characteristics, and t-SNE visualizations, are comprehensively presented in Fig. 5 and Table 1 . Table 1 Classification report of the MobileNetV2 model performance for each tree species, presenting key evaluation metrics and confusion matrix components. Precision Recall F1-score TP FP FN AUC Astronium fraxinifolium 0.96 0.77 0.86 55 2 16 0.96 Apuleia leiocarpa 0.80 1.00 0.89 72 18 0 0.98 Vochysia haenkeana 1.00 0.94 0.97 64 0 4 0.99 Accuracy 0.91 Macro AVG 0.92 0.91 0.91 Weighted AVG 0.92 0.91 0.91 * True Positives (TP); False Positives (FP); False Negatives (FN). The confusion matrix analysis (Table 1 ) revealed an overall prediction accuracy of 90.52%. Consistent with this, the weighted average F1-score was 0.91, indicating a good balance between precision and recall across all species. The highest classification error rate (22.54% of misclassifications, corresponding to 16 False Negatives for A . fraxinifolium and 18 False Positives for A. leiocarpa ) was observed between A . fraxinifolium and A. leiocarpa . This difficulty is associated with these species sharing common intrinsic bark characteristics such as rhytidome texture, rugosity, color patterns, and scars, whose variability is further compounded by environmental factors like the presence of lichens and bryophytes during the rainy season, which increase visual complexity. Detailed performance metrics for each species are also presented in Table 1 . Conversely, V. haenkeana showed superior classification performance, with a low error rate of 5.88%. This species achieved a perfect Precision of 1.00 and a high Recall of 0.94, resulting in an F1-score of 0.97. The discriminative ability for V. haenkeana was further supported by a high Area Under the Curve (AUC) value of 0.99. For A. fraxinifolium and A. leiocarpa, AUC values were also high at 0.96 and 0.98, respectively. Discussion The rapid advancements in convolutional neural networks (CNN’s) and deep learning techniques have profoundly transformed species identification, notably by enabling the analysis of complex morphological patterns (Kattenborn et al. 2021 ; Zhang et al. 2022 ). Historically, primary diagnostic features for plant identification have centered on leaves (Quach et al. 2023 ), flowers and fruits (Wang et al. 2022 ). However, recent studies have increasingly highlighted the critical role of bark texture as a robust alternative or complementary tool, particularly in contexts where phenological traits are seasonally absent or unreliable(Bhusnurmath and Doddamani 2023; Kim et al. 2022 ). This progress is especially pertinent in biodiverse tropical regions like the Cerrado, where high species coexistence and pronounced phenological variations, influenced by environmental factors such as seasonality, climate, wildfires, and anthropogenic activities, complicate traditional identification efforts (Katsis et al. 2022 ). Our findings align with this growing recognition, demonstrating the viability of bark morphology as a distinguishing characteristic for species classification in such complex environments. The findings of this study strongly support our first research question, confirming that species’ inherent morphological complexity directly influences CNN classification accuracy. While Fig. 1 indeed illustrates a range of bark morphologies among our selected species, from highly distinct to subtly similar patterns, this was a deliberate and integral aspect of our experimental design, not a fragility. This strategic selection allowed us to rigorously evaluate the evaluate the model's performance across a spectrum of bark morphological distinctiveness and to gain valuable insights into the specific challenges CNN’s face when differentiating species based on bark. As hypothesized, V. haenkeana , characterized by its more distinct and uniform textural patterns (Fig. 4 ), achieved superior performance, including a precision rate of 100% and a recall of 94%. This high performance is likely attributed to V. haenkeana smoother bark, generally lacking marked textures like prominent lenticels or fissures (Fig. 1 A-D). Conversely, A. fraxinifolium and A. leiocarpa exhibited higher misclassification rates due to their significant morphological similarity, particularly in shared features such as lenticels and scars, which present a more intricate and often convergent appearance. Specifically, A. leiocarpa bark is typically fissured, rugose, and exfoliates rounded irregular plates, varying in color from brown to gray, whitish, yellowish, or reddish (Fig. 1 E-H). Similarly, A. fraxinifolium bark is fissured, rugose, and exfoliating in rounded irregular plates, with a color range from brown to gray, whitish, yellowish, or reddish (Fig. 1 I-L). These results underscore the critical role of textural morphology in species differentiation by CNN’s, consistent with previous studies (Warner et al. 2025 ; Warner et al. 2024b ). While bark characteristics are known to vary with ontogeny (e.g., thickness and rugosity changing with tree age/size), our sampling approach, which focused on mature individuals with a minimum DBH of 25 cm, aimed to standardize the developmental stage of the sampled trees. However, residual intra-specific variation related to ontogeny could still influence classification performance and represent an inherent complexity of natural datasets. Our results further confirm the significant influence of common surface elements and environmental variability on the model's classification performance, directly addressing our second research question. The observed greater difficulty in classifying A. fraxinifolium and A. leiocarpa is notably linked to their shared superficial characteristics, such as lenticels, protuberances, and scars (Fig. 5 C). This challenge is exacerbated by the seasonal abundance of epiphytes like lichens and bryophytes during the rainy season, which introduce substantial visual noise and confounding features (Fig. 5 C). These elements can obscure subtle bark patterns, effectively blurring the visual distinctions between species, thus hindering CNN's ability to extract precise discriminatory features. This highlights a critical challenge for automated identification systems operating in dynamic natural environments, where visual heterogeneity is inherent and fluctuates with environmental conditions. In comparison to previous studies, this model demonstrates notable performance, achieving 91% overall accuracy in a highly biodiverse ecosystem like the Cerrado, which surpasses the average accuracy of 89% reported in similar CNN-based tree species identification efforts (Liu et al. 2019 ). This high performance underscores the practical applicability of our model in supporting large-scale forest inventories and environmental monitoring initiatives. Specifically, the model's capacity to rapidly and accurately assimilate species with simpler textural patterns, as observed for V. haenkeana , suggests its immediate utility for initial species screening in extensive projects. Indeed, our findings corroborate the significance potential of AI-based tools to enhance accuracy in forest inventories and support decision-making in tropical forest management (Gama et al. 2025 ). Despite the promising performance, this study acknowledges inherent limitations that offer avenues for future research. Firstly, while focusing on mature trees with a standardized DBH minimized ontogenetic variation, comprehensive assessment of temporal patterns and a broader representation of environmental variability (e.g., light conditions, canopy openness across seasons beyond just rainy season) were not exhaustively captured in the current dataset. Secondly, the model's reliance on bark images alone, while advantageous for phenology-independent identification, means it does not integrate other potentially diagnostic features (e.g., leaf, flower, fruit morphology, or phylogenetic context), which could further enhance accuracy, particularly for species with highly convergent bark patterns. To overcome these limitations and further advance the application of AI in ecological research, future work should prioritize expanding dataset size and diversity to include multiple seasons and broader geographical ranges. Additionally, integrating human expertise for iterative model refinement is crucial. Exploring multimodal deep learning approaches that combine various plant features (e.g., bark, leaf, flower, fruit morphology, or phylogenetic context) could significantly enhance accuracy, applicability, and robustness in diverse and dynamic ecosystems such as the Cerrado(Lapkovskis et al. 2024 ; Yang et al. 2025b ). This is a pertinent investigation, as traditional dichotomous keys for family and genus identification are primarily based on reproductive organs, with limited indications that bark patterns (rhytidoma) could be reliably used. However, the aid of AI, particularly in macro analysis, presents a promising avenue to investigate if structural rhytidome patterns and textural patterns are linked to families and genera, thereby expanding the utility of bark-based identification. Conclusion This study demonstrates the strong potential of the MobileNetV2 model for tree species identification in the Brazilian Cerrado using bark images, achieving an overall accuracy of 90.52%. The results confirm that bark morphological complexity plays a key role in classification performance: species with distinctive bark patterns, such as Vochysia haenkeana , were classified more accurately, whereas species with convergent characteristics, including Astronium fraxinifolium and Apuleia leiocarpa , posed greater challenges. The presence of superficial elements (e.g., lichens and scars) and environmental variability also affected model performance by introducing visual noise. Despite these challenges, MobileNetV2 represents a valuable complementary tool for large-scale forest inventories and environmental monitoring. Its phenology-independent approach is particularly advantageous in seasonally dynamic biomes like the Cerrado, especially where botanical expertise is limited, contributing to more efficient biodiversity assessments and conservation planning. However, this study has limitations, notably the use of data from a single seasonal period and the focus on species-level classification, highlighting the need for broader temporal coverage and taxonomic scope. Future research should expand the dataset across seasons and regions, integrate expert knowledge for iterative model refinement, and explore multimodal deep learning approaches combining multiple plant traits. Such advances could enhance accuracy, robustness, and the potential to extend bark-based identification beyond the species level, reinforcing the role of AI in biodiversity conservation. Declarations Acknowledgements JS Reis would like to thank CAPES (Coordenação de Aperfeiçoamento de Pessoal de Nível Superior) for the scholarship. CM Silva-Neto thanks CNPq (Conselho Nacional de Desenvolvimento Científico e Tecnológico) for the productivity grant. E Tizo-Pedroso thanks CNPq (408977/2016-7), Fundação de Amparo à Pesquisa do Estado de Goiás (FAPEG) (201610267001020), and Universidade Estadual de Goiás (Programa Pró-Pesquisa, 202200020022764) (Projetos Estratégicos, 202400020016002) for funding grants. Funding Funding was provided to E. Tizo-Pedroso by Universidade Estadual de Goiás (Programa Pró-Pesquisa, 202200020022764) (Projetos Estratégicos, 202400020016002), by Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq, 408977/2016-7), and ), Fundação de Amparo à Pesquisa do Estado de Goiás (FAPEG) (201610267001020). CRediT authorship contribution statement J.S.R: conceptualization, methodology, visualization, validation, data curation, formal analysis, writing-original draft. M.H.A.D.: Data curation and Formal analysis. C.M.S.N. and E.T.P.: supervision, resources, validation, writing-review and editing. All authors reviewed the manuscript. Declaration of Interest Statement The authors declare no competing interests. Data availability statement The datasets generated during and/or analyzed during the conduction of the study are available from the corresponding author upon reasonable request. References Aggarwal CC (2018) Convolutional Neural Networks. In: Aggarwal CC (ed) Neural Networks and Deep Learning: A Textbook. Springer International Publishing, Cham, pp 315–371. doi: https://doi.org/10.1007/978-3-319-94463-0_8 Akay M et al (2021) Deep Learning Classification of Systemic Sclerosis Skin Using the MobileNetV2 Model. IEEE Open J Eng Med Biology 2:104–110. 10.1109/OJEMB.2021.3066097 Al-Gaashani MSAM, Xu W, Obsie EY (2025) MobileNetV2-based deep learning architecture with progressive transfer learning for accurate monkeypox detection. Appl Soft Comput 169:112553. https://doi.org/10.1016/j.asoc.2024.112553 Altal OF et al (2025) Hybrid attention-enhanced MobileNetV2 with particle swarm optimization for endometrial cancer classification in CT images Informatics. Med Unlocked 57:101662. https://doi.org/10.1016/j.imu.2025.101662 Anselin L, Syabri I, Kho Y (2005) Anal Geographical Anal 38:5–22. 10.1111/j.0016-7363.2005.00671.x . GeoDa: An Introduction to Spatial Data Arora V, Rawat D, Singh N, Dwivedi SP, Chaturvedi P (2025) Leaf Based Plant Identification Using Traditional and Automated Methods. Rev Plant Archives 25:495–509. https://doi.org/10.51470/PLANTARCHIVES.2025.v25.supplement-1.067 Baihaqi MN, Sugiharto A, Tantyoko H (2025) Classification of Real and Fake Images Using Error Level Analysis Technique and MobileNetV2 Architecture. Jurnal Masyarakat Informatika 16:54–68. https://doi.org/10.14710/jmasif.16.1.73283 Berte ML, Quirino CB, Françoso R (2024) The economic potential of non-timber forest products in the Cerrado legal reserves - El potencial econômico de los productos forestales no madereros en las reservas legales de. El Cerrado Caldasia 46:381–394. 10.15446/caldasia.v46n2.102712 Bhusnurmath RA, Doddamani S Bark Texture Classification Using Deep Transfer Learning. In: Morusupalli R, Dandibhotla TS, Atluri VV, Windridge D, Lingras P, Komati VR (eds) Multi-disciplinary Trends in Artificial Intelligence, Cham, 2023// 2023. Springer Nature Switzerland, pp 407–420. https://doi.org/10.1007/978-3-031-36402-0_38 Blaanco LJ, Travieso CM, Quinteiro JM, Hernandez PV, Dutta MK (2016) Singh A A bark recognition algorithm for plant classification using a least square support vector machine. In: Ninth International Conference on Contemporary Computing (IC3), 11–13 Aug. 2016 2016. pp 1–5. 10.1109/IC3.2016.7880233 Caccianiga M, Compostella C, Caccia G, Cattaneo C (2021) Contribution of plant anatomy to forensic investigation: Tree bark morphology. Forensic Sci Int 318:110598. https://doi.org/10.1016/j.forsciint.2020.110598 Cicero M, Bilbily A, Colak E, Dowdell T, Gray B, Perampaladas K, Barfett J (2017) Training and Validating a Deep Convolutional Neural Network for Computer-Aided Detection and Classification of Abnormalities on Frontal Chest. Radiographs Invest Radiol 52. 10.1097/RLI.0000000000000341 Colli GR, Vieira CR, Dianese JC (2020) Biodiversity and conservation of the Cerrado: recent advances and old challenges. Biodivers Conserv 29:1465–1475. 10.1007/s10531-020-01967-x Elgamily KM, Mohamed MA, Abou-Taleb AM, Ata MM (2025) A novel W13 deep CNN structure for improved semantic segmentation of multiple objects in remote sensing imagery Neural. Comput Appl 37:5397–5427. 10.1007/s00521-024-10765-3 Enríquez-de-Salamanca Á Botanical databases in EIA: opportunities and challenges Impact Assessment and Project Appraisal:1–11 10.1080/14615517.2025.2482137 Fan C-L (2025) Evaluation Model for Crack Detection with Deep Learning: Improved Confusion Matrix Based on Linear Features. J Constr Eng Manag 151:04024210. 10.1061/JCEMD4.COENG-14976 Fekri-Ershad S (2020) Bark texture classification using improved local ternary patterns and multilayer neural network Expert Systems with Applications 158:113509 https://doi.org/10.1016/j.eswa.2020.113509 Gama NC et al (2025) Automatic Recognition of Commercial Tree Species from the Amazon Flora Using Bark Images and Transfer Learning. Preprints. 10.20944/preprints202506.1393.v1 Gulzar Y (2023) Fruit Image Classification Model Based on MobileNetV2 with Deep Transfer Learning Technique. Sustainability 15:1906. 10.3390/su15031906 Hajjaji Y, Boulila W, Farah IR, Koubaa A (2025) Enhancing palm precision agriculture: An approach based on deep learning and UAVs for efficient palm tree detection. Ecol Inf 85:102952. https://doi.org/10.1016/j.ecoinf.2024.102952 Hortal J, De-Bello F, Diniz-Filho JAF, Lewinsohn TM, Lobo JM, Ladle RJ (2015) Seven shortfalls that beset large-scale knowledge of biodiversity Annual review of ecology. Evol Syst 46:523–549. 10.1146/annurev-ecolsys-112414-054400 Huo R, Glickman ME (2025) Receiver operating characteristic analysis for paired comparison data. J Royal Stat Soc Ser A: Stat Soc 188:741–764. 10.1093/jrsssa/qnae072 Indraswari R, Rokhana R, Herulambang W (2022) Melanoma image classification based on MobileNetV2 network. Procedia Comput Sci 197:198–207. https://doi.org/10.1016/j.procs.2021.12.132 Katsis LKD, Hill AP, Piña-Covarrubias E, Prince P, Rogers A, Patrick Doncaster C, Snaddon JL (2022) Automated detection of gunshots in tropical forests using convolutional neural networks. Ecol Ind 141:109128. https://doi.org/10.1016/j.ecolind.2022.109128 Kattenborn T, Leitloff J, Schiefer F, Hinz S (2021) Review on Convolutional Neural Networks (CNN) in vegetation remote sensing ISPRS. J Photogrammetry Remote Sens 173:24–49. https://doi.org/10.1016/j.isprsjprs.2020.12.010 Kim TK et al (2022) Identifying and extracting bark key features of 42 tree species using convolutional neural networks and class activation mapping Scientific reports 12:4772 https://doi.org/10.1038/s41598-022-08571-9 Kumar T, Brennan R, Mileo A, Bendechache M (2024) Image Data Augmentation Approaches: A Comprehensive Survey and Future. Dir IEEE Access 4:1–1. 10.1109/ACCESS.2024.3470122 Kumaresan S, Aultrin KSJ, Kumar SS, Anand MD (2021) Transfer Learning With CNN for Classification of Weld. Defect IEEE Access 9:95097–95108. 10.1109/ACCESS.2021.3093487 Lapkovskis A, Nefedova N, Beikmohammadi A (2024) Automatic Fused Multimodal Deep Learning for Plant Identification arXiv preprint arXiv. 240601455:1–26. https://doi.org/10.48550/arXiv.2406.01455 Leiva-Bianchi M, Castillo N, Astudillo CA, Ahumada-Méndez F (2025) Effectiveness of machine learning methods in detecting grooming: a systematic meta-analytic. Rev Sci Rep 15:9008. 10.1038/s41598-024-83003-4 Li Z, Yang W, Peng S, Liu F (2020) A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects. 10.48550/arXiv.2004.02806 Lira-Martins D et al (2022) Soil properties and geomorphic processes influence vegetation composition, structure, and function in the Cerrado. Domain Plant Soil 476:549–588. https://doi.org/10.1007/s11104-022-05517-y Liu J, Wang X, Wang T (2019) Classification of tree species and stock volume estimation in ground forest images using Deep. Learn Computers Electron Agric 166:105012. https://doi.org/10.1016/j.compag.2019.105012 Liu J, Xing T, Wang X (2025) Optimizing AlexNet for accurate tree species classification via multi-branch architecture and mixed-domain attention. Complex Intell Syst 11:244. 10.1007/s40747-025-01836-6 Martello F et al (2023) Landscape structure shapes the diversity of plant reproductive traits in agricultural landscapes in the Brazilian Cerrado Agriculture. Ecosyst Environ 341:108216. https://doi.org/10.1016/j.agee.2022.108216 Meyer C, Weigelt P, Kreft H (2016) Multidimensional biases, gaps and uncertainties in global plant occurrence information. Ecol Lett 19:992–1006. https://doi.org/10.1111/ele.12624 Miftahushudur T, Sahin HM, Grieve B, Yin H (2025) A survey of methods for addressing imbalance data problems in. Agric Appl Remote Sens 17:454. https://doi.org/10.3390/rs17030454 Misra D, Crispim-Junior C, Tougne L, Patch-Based CNN Evaluation for Bark Classification. In: Bartoli A, Fusiello A (eds) Computer Vision – ECCV 2020 Workshops, Cham, 2020// 2020. Springer International Publishing, pp 197–212. https://doi.org/10.1007/978-3-030-65414-6_15 Mohammad AA, Dara R, Sunil K (2025) 7 - t-Distributed stochastic neighbor embedding. In: Amani Rad J, Chakraverty S, Parand K (eds) Dimensionality Reduction in Machine Learning. Morgan Kaufmann, pp 187–207. https://doi.org/10.1016/B978-0-44-332818-3.00017-4 Moran M, Faria M, Giraldi G, Bastos L, Oliveira L, Conci A (2021) Classification of Approximal Caries in Bitewing Radiographs Using Convolutional Neural Networks. Sensors 21. 10.3390/s21155192 Mustin K et al (2017) Biodiversity, threats and conservation challenges in the Cerrado of Amapá, an Amazonian. savanna Nat Conserv 22:107–127. 10.3897/natureconservation.22.13823 Nogueira K, Santos JA, Menini N, Silva TSF, Morellato LPC, Torres RS (2019) Spatio-Temporal Vegetation Pixel Classification by Using Convolutional Networks IEEE Geoscience and. Remote Sens Lett 16:1665–1669. 10.1109/LGRS.2019.2903194 Nooralishahi P et al (2021) Drone-Based Non-Destructive Inspection of Industrial Sites: A Review and Case Studies. Drones 5. https://doi.org/10.3390/drones5040106 QGIS (2022) Development Team QGIS Geographic Information System. Open Source Geospatial Foundation Project. http://qgis.osgeo.org . Accessed 12 September 2022 Quach BM, Dinh VC, Pham N, Huynh D, Nguyen BT (2023) Leaf recognition using convolutional neural networks based features Multimedia. Tools Appl 82:777–801. 10.1007/s11042-022-13199-y Rabeling SC, Lim JL, Tidon R, Neff JL, Simpson BB, Pawar S (2019) Seasonal variation of a plant-pollinator network in the Brazilian Cerrado: Implications for community structure and robustness. PLoS ONE 14:e0224997. https://doi.org/10.1371/journal.pone.0224997 Rocha LGS, Pinto JRR (2021) Diversity and structural changes in Cerrado Rupestre under effects of disturbances Brazilian. J Bot 44:213–225. https://doi.org/10.1007/s40415-020-00687-4 Rosell JA, Gleason S, Méndez-Alonzo R, Chang Y, Westoby M (2014) Bark functional ecology: evidence for tradeoffs, functional coordination, and environment producing bark diversity. New Phytol 201:486–497. https://doi.org/10.1111/nph.12541 Rouhan G, Gaudeul M (2014) Plant Taxonomy: A Historical Perspective, Current Challenges, and Perspectives. In: Besse P (ed) Molecular Plant Taxonomy: Methods and Protocols. Humana, Totowa, NJ, pp 1–37. doi: 10.1007/978-1-62703-767-9_1 Rzanny M, Mäder P, Deggelmann A, Chen M, Wäldchen J (2019) Flowers, leaves or both? How to obtain suitable images for automated plant identification. Plant Methods 15:77. 10.1186/s13007-019-0462-4 Schwaida SF, Cicerelli RE, Almeida T, Sano EE, Pires CH, Ramos APM (2023) Defining priorities areas for biodiversity conservation and trading forest certificates in the Cerrado biome in. Brazil Biodivers Conserv 32:1807–1820. https://doi.org/10.1007/s10531-023-02578-y Shah P, Mishra DK, Shanmugam MP, Doshi B, Jayaraj H, Ramanjulu R (2020) Validation of Deep Convolutional Neural Network-based algorithm for detection of diabetic retinopathy – Artificial intelligence versus clinician for screening. Indian J Ophthalmol 68. 10.4103/ijo.IJO_966_19 Shorten C, Khoshgoftaar TM, Furht B (2021) Text Data Augmentation for Deep Learning. J Big Data 8:101. 10.1186/s40537-021-00492-0 Silva-Pereira BA, Venturoli F, Carvalho FA (2011) Florestas Estacionais no Cerrado: Uma Visão. Geral Pesquisa Agropecuária Trop 446–455. https://doi.org/10.5216/pat.v41i3.12666 Silveira P et al (2025) The loss of an unknown biodiversity: Spatial gaps in plant survey and conservation in a Brazilian. hotspot Biodivers Biol Conserv 305:111098. https://doi.org/10.1016/j.biocon.2025.111098 Souza KR, Oliveira G, Ferreira F, Martins TO, Martins T, Venturoli F, Silva-Neto C (2016a) Floresta Estacional: madeira para o produtor rural no Cerrado Seasonal Forest: Sustainable wood for farmers in. Brazilian Savanna Magistra 28:548–556 Souza KR, Oliveira GM, Ferreira FG, Martins TO, Melo C, Neto S, Venturoli F (2016b) Floresta Estacional: madeira sustentável para o produtor. rural no Cerrado MAGISTRA 28:54–62 Surendran GK, Deekshitha, Lukac M, Lukac M, Vybostok J, Mokros M (2025) A forestry investigation: Exploring factors behind improved tree species classification using bark images. Ecol Inf 85:102932. https://doi.org/10.1016/j.ecoinf.2024.102932 Wang C et al (2022) Application of Convolutional Neural Network-Based Detection Methods in Fresh Fruit Production: A Comprehensive Review. Front Plant Sci 13. 10.3389/fpls.2022.868745 Warner C, Wu F, Gazo R, Benes B, Fei S (2025) Environmental Sensitivity in AI Tree Bark Detection: Identifying Key Factors for Improving Classification Accuracy. Algorithms 18. 10.3390/a18070417 Warner C, Wu F, Gazo R, Benes B, Kong N, Fei S (2024a) Central Bark Image Dataset and Tree Species Classification Using Deep Learning. Algorithms 17. https://doi.org/10.3390/a17050179 Warner C, Wu F, Gazo R, Benes B, Kong N, Fei S (2024b) CentralBark Image Dataset and Tree Species Classification Using Deep Learning. Algorithms 17. 10.3390/a17050179 Wu F, Gazo R, Benes B, Haviarova E (2021) Deep BarkID: a portable tree bark identification system by knowledge distillation European. J For Res 140:1391–1399. https://doi.org/10.1007/s10342-021-01407-7 Xu Y, Goodacre R (2018) On Splitting Training and Validation Set: A Comparative Study of Cross-Validation, Bootstrap and Systematic Sampling for Estimating the Generalization Performance of Supervised Learning. J Anal Test 2:249–262. 10.1007/s41664-018-0068-2 Yang G, Wang G, Wan L, Wang X, He Y (2025a) Utilizing SMOTE-TomekLink and machine learning to construct a predictive model for elderly medical and daily care services demand. Sci Rep 15:8446. 10.1038/s41598-025-92722-1 Yang Z-X, Li Y, Wang R-F, Hu P, Su W-H (2025b) Deep Learning in Multimodal Fusion for Sustainable Plant Care: A Comprehensive Review Sustainability (2071 – 1050). 17. 10.3390/su17125255 Zhang Y, Wang Z, Lu Y, Zuo L (2022) Biodiversity, ecosystem functions and services: Interrelationship with environmental and human health. Front Ecol Evol 10:1086408. https://doi.org/10.3389/fevo.2022.1086408 Zhang Z, Ao Z, Wu W, Wang Y, Xin Q (2024) Developing a Multi-Scale Convolutional Neural Network for Spatiotemporal Fusion to Generate MODIS-like Data Using AVHRR and Landsat Images. Remote Sens 16. https://doi.org/10.3390/rs16061086 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8841518","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":591925611,"identity":"92514477-e602-4100-b7fe-27fe672af099","order_by":0,"name":"Jéssica Silva Reis","email":"","orcid":"","institution":"State University of Goiás","correspondingAuthor":false,"prefix":"","firstName":"Jéssica","middleName":"Silva","lastName":"Reis","suffix":""},{"id":591925612,"identity":"3fe4c54d-d767-4f10-96e6-7face146d766","order_by":1,"name":"Marcos Henrique André Deus","email":"","orcid":"","institution":"State Military Police School Dr. Belém. Bela Vista. Goiás","correspondingAuthor":false,"prefix":"","firstName":"Marcos","middleName":"Henrique André","lastName":"Deus","suffix":""},{"id":591925613,"identity":"055bdd04-73eb-49e3-b56f-a40673a2126d","order_by":2,"name":"Carlos Silva-Neto","email":"","orcid":"","institution":"Federal Institute of Goiás, State University of Goiás","correspondingAuthor":false,"prefix":"","firstName":"Carlos","middleName":"","lastName":"Silva-Neto","suffix":""},{"id":591925614,"identity":"01648290-ad6f-475a-b36f-d6abeb38ffd9","order_by":3,"name":"Everton Tizo-Pedroso","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzElEQVRIiWNgGAWjYDACZgTJ+ABEGhDSwYOkhdmAOC1IdrFJEKXFnp394eOCCut8fv7Dx6p5KmoZzKUPEHRYsvGMM+mWM2ekpd3mOXOcwbIvgaCWY9K8bYcNDG7wmN3mbTvGYHCGkF+YGdukef8BtZw/Y1bM+48oLcxs0rwNQC0HcsyYeRtqiNBymI3ZeMaxdAPJGWnJknOOHeCx7CGghb3/ODDEaqwNgCF28MObmjo5cx4CWkCAGYl9mBgNqFrqiNIxCkbBKBgFIwsAABcmOGCHFEmTAAAAAElFTkSuQmCC","orcid":"","institution":"State University of Goiás","correspondingAuthor":true,"prefix":"","firstName":"Everton","middleName":"","lastName":"Tizo-Pedroso","suffix":""}],"badges":[],"createdAt":"2026-02-10 13:24:58","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8841518/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8841518/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102940151,"identity":"1c3950cf-42ce-4c52-bf31-2bbbac5f16fb","added_by":"auto","created_at":"2026-02-18 17:04:15","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":540890,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative examples of bark morphological variation within species and similarities between species. (A-D) Bark of \u003cem\u003eVochysia haenkeana \u003c/em\u003e(Vochysiaceae), illustrating intraspecific variation in texture and coloration. (E-H) Bark of \u003cem\u003eApuleia leiocarpa \u003c/em\u003e(Fabaceae), illustrating intraspecific variation in texture and coloration. (I-L) Bark of \u003cem\u003eAstronium fraxinifolium\u003c/em\u003e (Anacardiaceae), illustrating intraspecific variation in texture and coloration. These detailed views highlight the complex morphological variations within each species and underscore the subtle similarities present between \u003cem\u003eVochysia haenkeana\u003c/em\u003e, \u003cem\u003eApuleia leiocarpa\u003c/em\u003e and \u003cem\u003eAstronium fraxinifolium\u003c/em\u003e (e.g., compare panels F and J, or G and K), which contribute to identification challenges\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8841518/v1/cf20eccf480e51b3498ab17b.jpeg"},{"id":102940153,"identity":"6cbada45-0d40-4a83-b541-f0b75e992880","added_by":"auto","created_at":"2026-02-18 17:04:15","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1648287,"visible":true,"origin":"","legend":"\u003cp\u003eSpatial distribution of sampled tree individuals in the Brazilian Cerrado biome across nested geographical contexts. (A) Location of Goiás State within Brazil. (B) Overview of Goiás State. (C) Detailed map of sampled individuals in Abadia de Goiás municipality on satellite imagery. Points are color-coded by species: blue diamonds for \u003cem\u003eApuleia leiocarpa\u003c/em\u003e, red circles for \u003cem\u003eAstronium fraxinifolium\u003c/em\u003e, and yellow triangles for \u003cem\u003eVochysia haenkeana\u003c/em\u003e. (D) Detailed map of sampled individuals in Piracanjuba municipality on satellite imagery. Points are color-coded by species: blue diamonds for \u003cem\u003eApuleia leiocarpa\u003c/em\u003e, red circles for \u003cem\u003eAstronium fraxinifolium\u003c/em\u003e, and yellow triangles for \u003cem\u003eVochysia haenkeana\u003c/em\u003e. The map was prepared by JS Reis using the software QGIS v.3.38.2 \u0026nbsp;(QGIS 2022) and the Landsate 8 satellite image.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8841518/v1/e6e92d961385f13d066f972c.png"},{"id":102940152,"identity":"fb2212c1-7282-4ab7-b104-62e3f00729f5","added_by":"auto","created_at":"2026-02-18 17:04:15","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":339695,"visible":true,"origin":"","legend":"\u003cp\u003eOverall methodological pipeline for bark image classification, illustrating the sequential workflow from image collection and preprocessing to model training, evaluation, and interpretation.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8841518/v1/dcf856abfc69e4806ad0047a.png"},{"id":102940154,"identity":"475bf880-83aa-4e72-95d2-b498af1b63ea","added_by":"auto","created_at":"2026-02-18 17:04:15","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":364119,"visible":true,"origin":"","legend":"\u003cp\u003eQualitative assessment of key bark morphological characteristics for the three tree species: \u003cem\u003eAstronium fraxinifolium\u003c/em\u003e, \u003cem\u003eApuleia leiocarpa\u003c/em\u003e, and \u003cem\u003eVochysia haenkeana\u003c/em\u003e. The analyzed dimensions include bark texture, bark thickness, shedding pattern, outer coloration, color change with shedding, and the presence of surface elements. These characteristics were visually assessed to represent the inherent morphological variability and distinctiveness among the species, which directly influenced the model's learning and differentiation challenges.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8841518/v1/88412132d35a96dab1f6d14f.png"},{"id":102940156,"identity":"046d3651-7917-41e4-8f6a-a095944af73a","added_by":"auto","created_at":"2026-02-18 17:04:16","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":270592,"visible":true,"origin":"","legend":"\u003cp\u003eComprehensive visualization of MobileNetV2 model performance and underlying bark characteristics for species classification. (A) Confusion Matrix: Displays the classification performance, showing true labels versus predicted labels for \u003cem\u003eApuleia leiocarpa\u003c/em\u003e, \u003cem\u003eAstronium fraxinifolium\u003c/em\u003e, and \u003cem\u003eVochysia haenkeana\u003c/em\u003e. (B) Receiver Operating Characteristic (ROC) Curves: Illustrates the discriminative capabilities of the model for each species, with corresponding Area Under the Curve (AUC) values. (C) Heatmap of Bark Characteristics: Shows the distribution and visually assessed scores of key bark morphological characteristics across the three species. The numerical scale (1-3) represents a qualitative categorization of each bark trait's prominence or complexity, where 1 indicates low, 2 moderate, and 3 high prominence/complexity. This visualization highlights the inherent variability and similarities in bark traits influencing classification. (D) t-SNE Visualization: Displays the clustering (separation and overlap) of specific patterns extracted from the image samples in a two-dimensional space. Points represent 1,304 samples, color-coded by species: \u003cem\u003eA. fraxinifolium\u003c/em\u003e(purple/blue), \u003cem\u003eA. leiocarpa\u003c/em\u003e (green), and \u003cem\u003eV. haenkeana\u003c/em\u003e (yellow). The x-axis represents t-SNE feature 1, and the y-axis represents t-SNE feature 2. This panel visually demonstrates the learned separability and inherent similarities among the species' bark features.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8841518/v1/f5c765eff160845c6271cf11.png"},{"id":103049614,"identity":"1a15fa99-1cf1-4937-8380-549c175678d4","added_by":"auto","created_at":"2026-02-20 07:43:54","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4472149,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8841518/v1/aaef0bca-942f-4a8b-bead-b8ae049c2b7e.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eOvercoming Structural and Environmental Challenges in Identifying Brazilian Savanna Tree Species Using Deep Learning\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eSouth American savannas represent a significant portion of global savanna coverage, spanning over 2.29\u0026nbsp;million square kilometers and distinguishing themselves as the wettest and most biodiverse spectrum globally (Lira-Martins et al. \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Rocha and Pinto \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Schwaida et al. \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). The Brazilian Cerrado, a globally recognized biodiversity hotspot, harbors an extraordinary diversity of endemic species essential for ecosystem functioning and resilience (Lira-Martins et al. \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Rabeling et al. \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Its structural and compositional complexity creates a mosaic of ecological niches, supporting exceptionally high levels of functional diversity (Lira-Martins et al. \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Rocha and Pinto \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Schwaida et al. \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Yet, despite its ecological prominence, the Cerrado remains critically understudied, posing substantial challenges for accurate species identification and evidence-based conservation planning (Colli et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Mustin et al. \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Silveira et al. \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Indeed, the comprehensive understanding of biodiversity is frequently hampered by \u0026ldquo;knowledge shortfalls\u0026rdquo;, where crucial research questions remain unanswered due to data deficiencies (Hortal et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). Such deficiencies are often exacerbated by the fact that traditional approaches, such as classical taxonomy and botanical field surveys, though foundational, face inherent challenges in providing comprehensive and efficient large-scale identification (Arora et al. \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Rzanny et al. \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). These challenges arise especially due to the psychophysiological complexity of seasonal forest formations and logistical constraints that hinder professional access to remote areas (Meyer et al. \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2016\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn response to these substantial identification challenges that hinder effective biodiversity management, novel technological solutions are rapidly transforming ecological assessment and monitoring. Among these, convolutional neural networks (CNN\u0026rsquo;s) have emerged as remarkably powerful and tools for large-scale species monitoring (Li et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Zhang et al. \u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Their unparalleled ability to analyze complex spatial, temporal, and spectral data with high precision stems from their design as deep learning models, capable of automatically identifying intricate hierarchical patterns - including edges, textures, and shapes - thereby circumventing the need for extensive manual feature engineering (Aggarwal \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Warner et al. \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e2024a\u003c/span\u003e). This inherent proficiency has driven their widespread adoption in diverse ecological applications, ranging from accurate species identification and habitat mapping to critical environmental monitoring, such as deforestation detection and assessment(Kim et al. \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Furthermore, CNN-based models have demonstrated significant potential for forest inventories, enabling rapid and accurate classification of plant species from remote sensing data and successfully integrated into drones and scanning systems for detailed insights into inaccessible forested areas (Nooralishahi et al. \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Wu et al. \u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eDespite the well-established efficacy of CNN\u0026rsquo;s in various Classification tasks, a significant research gap persists regarding their specific application for native Cerrado tree species. Identifying these species through bark morphology offers a non \u0026ndash; destructive approach, especially valuable when key phenological traits are seasonally inconspicuous (Kim et al. \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). This is crucial, as botanical surveys frequently encounter abundant sterile vegetative material, and species identification in the Cerrado is further complicated by intraspecific variation across diverse habitats (e.g., forests, open fields, various soil types) (Blaanco et al. \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Fekri-Ershad \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Furthermore, bark offers high consistency throughout seasons and easy accessibility, even in high crown conditions (Rosell et al. \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). However, its full potential for automated identification remains critically underexplored for species within semideciduous seasonal forest formations (Fekri-Ershad \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). This gap is critical given the unique complexities of this biome and the urgent need for enhanced conservation tools. While expert botanists can often distinguish these species based on a combination of traits, complementary technological tools, including AI-driven systems, are increasingly essential to support field identification efforts (Caccianiga et al. \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Enr\u0026iacute;quez-de-Salamanca). Such tools can provide rapid, large-scale assessment, crucial for enhancing existing botanical knowledge and aiding conservation amidst rapid savanna destruction and climate change. This aligns with the broader development of interactive identification keys and comprehensive digital databases.\u003c/p\u003e \u003cp\u003eThis inherent difficulty in identification stems from subtle morphological variations, similar bark coloration, and shared intrinsic bark features (e.g., rhytidome texture, rugosity, color patterns, and scars) among co-occurring species (Gama et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Surendran et al. \u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). This challenge is especially exacerbated during dry seasons and fires, when key phenological traits are less available for accurate diagnosis(Lira-Martins et al. \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Rabeling et al. \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Such challenges make specialized knowledge and precise identification tools limiting students in early stages of training and non-specialized individuals. For instance, environmental plasticity leads to confusion even for seemingly distinct species: \u003cem\u003eVochysia haenkeana\u003c/em\u003e, despite its marked yellow coloration, is often mistaken for \u003cem\u003eAlbizia niopoides\u003c/em\u003e by non-specialists. Similarly, \u003cem\u003eApuleia leiocarpa\u003c/em\u003e and \u003cem\u003eAstronium fraxinifolium\u003c/em\u003e present highly similar rhytidome patterns and variable coloration in different environments (e.g., more brownish in closed forest vs. typical ashen/dark yellow in open environments), leading to frequent misidentification. Furthermore, these three species \u0026ndash; \u003cem\u003eV. haenkeana\u003c/em\u003e, \u003cem\u003eA. leiocarpa\u003c/em\u003e, and \u003cem\u003eA. fraxinifolium\u003c/em\u003e - can be confused with each other by the local community, as their bark tends to become more yellowish and subtly smooth depending on the time of year and climatic conditions, intensifying the ambiguity in visual identification.\u003c/p\u003e \u003cp\u003eThis ambiguity directly impacts the sustainable use of the Cerrado rich flora. Many species, particularly those in semideciduous seasonal forest formations, possess high economic potential (e.g., valuable timber, edible fruits, medicinal properties, non-timber forest products) (Berte et al. \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Silva-Pereira et al. \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2011\u003c/span\u003e). However, their sustainable exploitation and valorization are severely hindered by this ongoing identification ambiguity, preventing adequate economic valuation and limiting community engagement in sustainable forest management (Souza et al. \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2016a\u003c/span\u003e). Consequently, this contributes to the undervaluation and displacement of native resources by expanding agricultural frontiers, further exacerbating biodiversity loss (Rouhan and Gaudeul \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). Given the proven social acceptance of CNN\u0026rsquo;s among farmers and increasing initiatives for savanna plants thorough these networks are highly promising (Hajjaji et al. \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Nogueira et al. \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). This allows producers to identify valuable species without confusion, fostering sustainable economic development and empowering local communities within the Cerrado.\u003c/p\u003e \u003cp\u003eTo directly address this pressing research and practical challenge, this study investigates the applicability of advanced deep learning techniques. Specifically, we evaluate the MobileNetV2 convolutional neural network for classification of three ecologically prominent Cerrado tree species - \u003cem\u003eApuleia leiocarpa\u003c/em\u003e (Fabaceae), \u003cem\u003eAstronium fraxinifolium\u003c/em\u003e (Anacardiaceae), and \u003cem\u003eVochysia haenkeana\u003c/em\u003e (Vochysiaceae) - based solely on bark images. These species were selected for their ecological significance, widespread occurrence, and the morphological complexity of their bark surfaces, which present a nuanced and realistic challenge for automated recognition systems, as highlighted by the field identification difficulties and their economic/ecological value.\u003c/p\u003e \u003cp\u003eOur research is guided by two central questions, which also represent key methodological challenges in this context: Firstly, we aim to understand how the inherent morphological complexity of bark influences CNN classification accuracy, particularly when differentiating between species exhibiting simpler versus more intricate patterns. Secondly, we investigate how the model\u0026rsquo;s performance is influenced by natural variability of intrinsic bark characteristics. This includes assessing the impact of features integral to itself \u0026ndash; such as rhytidome texture and rugosity, color patterns, and the presence of scars \u0026ndash; which can act as confounding variables for the classifier. By systematically exploring these questions, this study aims to promote CNN-based species identification methodologies and advance the application of deep learning in biodiversity conservation within conservation within complex ecosystems. Our findings will not only underscore the promising role of artificial intelligence in ecological research but also transparently delineate the practical limitations and highlight crucial future directions for developing highly accurate and robust CNN-based identification systems tailored for challenging ecological contexts.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eField Study and Data Collection\u003c/h2\u003e \u003cp\u003eThis study employed a dataset specifically acquired to investigate bark texture patterns of tree species in the Brazilian savanna (Cerrado). A total of 1,515 images were systematically collected, evenly distributed among three native tree species within this biome: \u003cem\u003eApuleia leiocarpa\u003c/em\u003e (Vogel) J.F. Macbr. (Fabaceae), \u003cem\u003eAstronium fraxinifolium\u003c/em\u003e Schott (Anacardiaceae), and \u003cem\u003eVochysia haenkeana\u003c/em\u003e Mart. (Vochysiaceae). Crucially, each species contributed 505 unique images, with each image representing a distinct individual tree, all with a Diameter at Breast Height (DBH) equal to or greater than 25 cm. Images were captured during the rainy seasons of 2023 and 2024, in mesophytic forests located in southern Goi\u0026aacute;s State, Brazil, specifically the municipalities of Piracanjuba (17\u0026deg;18'30.42\"S, 49\u0026deg;1'44.37\"W) and Abadia de Goi\u0026aacute;s (16\u0026deg;45'25.44\"S, 49\u0026deg;25'35.12\"W). Representative examples of the bark morphology for each species are presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, including both general trunks. A map illustrating the spatial distribution of the sampled individuals is provided in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The selection of these species was previously justified in the Introduction based on their ecological significance, widespread occurrence, and the morphological complexities of their bark surfaces, which present a nuanced challenge for automated identification systems. This selection strategy was integral to our experimental design, allowing us to evaluate the model\u0026rsquo;s performance across a spectrum of bark morphological distinctiveness, from more uniform to subtly convergent patterns.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFor image acquisition, two types of equipment were utilized to ensure diversity and quality in the records: a Nikon D90 camera (equipped with an 18\u0026ndash;105 mm lens, 4288 x 2848 pixels resolution), and the integrated camera of a Samsung Galaxy A71 smartphone (4000 x 3000 pixels resolution). All images were stored in JPEG format. Acknowledging the inherent variability of natural light incidence and the dynamic influence of canopy openness in understory environments, efforts were made to standardize image capture. During collection, the imaging device was consistently positioned perpendicular to the tree trunk, at distance ranging from 20 cm to 40 cm, depending on the conditions found in the field (Gama et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Image capture was performed under natural lighting conditions, specifically between 9:00 AM and 3:00 PM, to mitigate extreme variations in light intensity and shadow effects. The residual variations in lighting and other environmental factors inherent to field data were specifically addressed during the dataset preprocessing through robust data augmentation techniques (as detailed in Section Dataset Preparation and Preprocessing), designed to enhance the model\u0026rsquo;s resilience to diverse real-world conditions.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eMachine Learning Pipeline Overview\u003c/h3\u003e\n\u003cp\u003eThe overall methodology for bark image classification, encompassing data preparation, model training, and evaluation, is schematically presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. This pipeline was designed to rigorously test our research questions regarding the influence of bark morphology and superficial elements on CNN classification accuracy. The following subsections detail each component of this pipeline.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eDataset Preparation and Preprocessing\u003c/h3\u003e\n\u003cp\u003eUpon collection, all images were initially in JPEG format and organized using the Hierarchical Dataset Organization (Xu and Goodacre \u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). To optimize feature extraction and computational efficiency for CNN training, and to enhance the detection of local texture patterns, the original high-resolution bark images were segmented into multiple smaller sub-images, or patches (Altal et al. \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). From each original image, 60 random patches of 256 x 256 pixels were extracted, a number considered adequate according to previous studies utilizing between 9 and 80 patches (Elgamily et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Misra et al. 2020). This resulted in a total dataset of 90,900 patches for analysis. This patch-based approach is particularly beneficial for training convolutional neural networks as it increases dataset variability and focuses the model on fine-grained textural details. This patch-based approach is particularly advantageous for training convolutional neural networks, as it increases dataset variability and optimizes the model's focus on fine-grained textural details (Baihaqi et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe dataset was systematically divided into stratified training, validation, and test sets to ensure robust model development and unbiased evaluation (Liu et al. \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). For each species, 85% of the images were allocated for model training and validation, while the remaining 15% were exclusively reserved for the test set (Cicero et al. \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Shah et al. \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). This partitioning ensures that the model's generalization performance is evaluated on unseen data, mitigating potential \u003cem\u003eoverfitting(\u003c/em\u003eAltal et al. \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2025\u003c/span\u003e\u003cem\u003e)\u003c/em\u003e.\u003c/p\u003e \u003cp\u003eTo further enhance the model's ability to generalize across diverse real-world lighting and orientation conditions, and to effectively prevent overfitting, data augmentation techniques were extensively applied to the training set (Shorten et al. \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). These techniques included randomized adjustments such as rotation (\u0026plusmn;\u0026thinsp;20\u0026deg;), zoom (\u0026plusmn;\u0026thinsp;20%), horizontal flipping, and brightness modifications (\u0026plusmn;\u0026thinsp;15%). Prior to model training, all images underwent a crucial normalization process (Kumar et al. \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), where pixel values were rescaled to the [0, 1] range. This preprocessing step ensured data uniformity, mitigated potential illumination variances, and guaranteed compatibility with the MobileNetV2 model\u0026rsquo;s input requirements (Indraswari et al. \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e\n\u003ch3\u003eMachine Learning\u003c/h3\u003e\n\u003cp\u003eThe MobileNetV2 architecture, a highly efficient convolutional neural network, was selected for its proven capability in detecting relevant features in images, making it suitable for deployment on mobile and edge devices (Gulzar \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). This architecture, configured with \u003cem\u003etransfer learning\u003c/em\u003e techniques, leverages a model pre-trained on the extensive ImageNet database(Akay et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). This pre-training provides an initial set of robust recognition features(Al-Gaashani et al. \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), such as edges, textures, and shapes, which are then fine-tuned to the specific characteristics of our bark image dataset.\u003c/p\u003e \u003cp\u003eHyperparameter optimization was a critical step to enhance the model\u0026rsquo;s robustness and reliability (Kumaresan et al. \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Beyond standard parameters like learning rate, number of epochs, and batch size, the optimization process aimed to fine-tune the model's capacity to automatically extract and differentiate salient morphological patterns from bark images. This approach enabled the model to effectively adapt to and discern subtle morphological specificities and potential visual noise inherent in our dataset, directly addressing aspects related to our research questions concerning bark complexity and superficial elements(Altal et al. \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Qualitative characteristics of bark morphology, including external coloration, texture, thickness, shedding pattern, color change with shedding, and the presence of superficial elements (e.g., lichens, bryophytes, scars), were visually assessed during dataset preparation (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). This visual assessment provided crucial insights for experimental design and aided in diagnosing model performance and classification challenges.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eModel Evaluation and Statistical Analysis\u003c/h3\u003e\n\u003cp\u003ePrior to model evaluation, and to investigate potential spatial autocorrelation that could affect subsequent analyses, a Moran's I analysis was conducted for key intrinsic bark characteristics, including rhytidome texture, bark dimension (length, width, thickness, and weight), rugosity, color patterns, and scars. This analysis, performed using the Geoda software v.1.8.10 (Anselin et al. \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2005\u003c/span\u003e), revealed no significant spatial autocorrelation for these variables, suggesting their spatial independence (Moran et al. \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe MobileNetV2 classification model's performance was rigorously evaluated on an independent test set using standard metrics, providing insights crucial for addressing our research questions. These metrics included Accuracy (ACC), representing overall correct predictions; Precision (P), measuring the quality of positive predictions; Recall (R), quantifying the model's sensitivity to identify all positive instances; and the F1-score (F1), the harmonic mean of Precision and Recall, indicating a balanced performance. Their calculations are defined as:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:ACC=\\frac{TP+TN}{TP+TN+FP+FN}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:P=\\frac{TP}{TP+TN}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:R=\\frac{TP}{TP+FN}\\:$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:F1=2*\\frac{P*R}{P+R}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eWhere TP, TN, FP, and FN represent True Positives, True Negatives, False Positives, and False Negatives, respectively(Leiva-Bianchi et al. \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Yang et al. \u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e2025a\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe confusion matrix was employed to systematically visualize classification performance across species. This matrix was critical for assessing the model's ability to distinguish species with similar bark characteristics(Fan \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Miftahushudur et al. \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), particularly \u003cem\u003eA. fraxinifolium\u003c/em\u003e and \u003cem\u003eA. leiocarpa\u003c/em\u003e, thereby directly informing our investigation into how morphological complexity influences CNN classification accuracy. It also facilitated the identification of misclassification patterns potentially linked to visual noise from superficial elements, contributing to our understanding of the impact of such confounding features on model performance(Miftahushudur et al. \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFor further interpretation of the model\u0026rsquo;s discriminative capabilities, Receiver Operating Characteristic (ROC) curves and their Area Under the Curve (AUC) were generated for each species (Huo and Glickman \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Additionally, t-Distributed Stochastic Neighbor Embedding (t-SNE) was applied to visualize high-dimensional feature vectors in two dimensions, revealing data clusters and patterns of similarity or overlap among species in the latent space (Mohammad et al. \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). These visualizations provided qualitative insights into how the model differentiates, or struggles to differentiate, species based on bark morphology and the impact of visual noise, supporting the interpretation of our research findings.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003ePlant Material\u003c/h2\u003e \u003cp\u003eThe three native Cerrado tree species - \u003cem\u003eApuleia leiocarpa\u003c/em\u003e (Fabaceae), \u003cem\u003eAstronium fraxinifolium\u003c/em\u003e (Anacardiaceae), and \u003cem\u003eVochysia haenkeana\u003c/em\u003e (Vochysiaceae) - were selected due to their ecological significance, widespread occurrence, and high economic potential in semideciduous seasonal forest formations (Souza et al. \u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e2016b\u003c/span\u003e). Their identification presents significant challenges stemming from subtle morphological variations, variable bark coloration (influenced by environment and season), and shared superficial features (e.g., coloration, texture, and scars). This complexity leads to frequent misidentification by non-specialists, with other co-occurring species like \u003cem\u003eAlbizia niopoides\u003c/em\u003e and among themselves (\u003cem\u003eA. leiocarpa\u003c/em\u003e, \u003cem\u003eA. fraxinifolium\u003c/em\u003e, and \u003cem\u003eV. haenkeana\u003c/em\u003e) (Martello et al. \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). These inherent ambiguities severely hinder the sustainable exploitation and valorization of these valuable resources, ultimately contributing to the undervaluation and displacement of native resources by agricultural expansion(Martello et al. \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Consequently, these species serve as ideal test cases for evaluating AI-driven tools in addressing real-world identification bottlenecks in complex and economically vital ecosystems.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eThe MobileNetV2 model demonstrated strong performance in classifying the three Cerrado tree species from bark images. The model\u0026rsquo;s performance metrics and visual analytics, including the confusion matrix, Receiver Operating Characteristic (ROC) curves, a heatmap of bark characteristics, and t-SNE visualizations, are comprehensively presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e and Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eClassification report of the MobileNetV2 model performance for each tree species, presenting key evaluation metrics and confusion matrix components.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eFP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eFN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eAstronium fraxinifolium\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eApuleia leiocarpa\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eVochysia haenkeana\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMacro AVG\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWeighted AVG\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"8\"\u003e* True Positives (TP); False Positives (FP); False Negatives (FN).\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe confusion matrix analysis (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) revealed an overall prediction accuracy of 90.52%. Consistent with this, the weighted average F1-score was 0.91, indicating a good balance between precision and recall across all species. The highest classification error rate (22.54% of misclassifications, corresponding to 16 False Negatives for \u003cem\u003eA\u003c/em\u003e. \u003cem\u003efraxinifolium\u003c/em\u003e and 18 False Positives for \u003cem\u003eA. leiocarpa\u003c/em\u003e) was observed between \u003cem\u003eA\u003c/em\u003e. \u003cem\u003efraxinifolium\u003c/em\u003e and \u003cem\u003eA. leiocarpa\u003c/em\u003e. This difficulty is associated with these species sharing common intrinsic bark characteristics such as rhytidome texture, rugosity, color patterns, and scars, whose variability is further compounded by environmental factors like the presence of lichens and bryophytes during the rainy season, which increase visual complexity.\u003c/p\u003e \u003cp\u003eDetailed performance metrics for each species are also presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Conversely, \u003cem\u003eV. haenkeana\u003c/em\u003e showed superior classification performance, with a low error rate of 5.88%. This species achieved a perfect Precision of 1.00 and a high Recall of 0.94, resulting in an F1-score of 0.97. The discriminative ability for \u003cem\u003eV. haenkeana\u003c/em\u003e was further supported by a high Area Under the Curve (AUC) value of 0.99. For A. fraxinifolium and A. leiocarpa, AUC values were also high at 0.96 and 0.98, respectively.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe rapid advancements in convolutional neural networks (CNN\u0026rsquo;s) and deep learning techniques have profoundly transformed species identification, notably by enabling the analysis of complex morphological patterns (Kattenborn et al. \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Zhang et al. \u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Historically, primary diagnostic features for plant identification have centered on leaves (Quach et al. \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), flowers and fruits (Wang et al. \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). However, recent studies have increasingly highlighted the critical role of bark texture as a robust alternative or complementary tool, particularly in contexts where phenological traits are seasonally absent or unreliable(Bhusnurmath and Doddamani 2023; Kim et al. \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). This progress is especially pertinent in biodiverse tropical regions like the Cerrado, where high species coexistence and pronounced phenological variations, influenced by environmental factors such as seasonality, climate, wildfires, and anthropogenic activities, complicate traditional identification efforts (Katsis et al. \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Our findings align with this growing recognition, demonstrating the viability of bark morphology as a distinguishing characteristic for species classification in such complex environments.\u003c/p\u003e \u003cp\u003eThe findings of this study strongly support our first research question, confirming that species\u0026rsquo; inherent morphological complexity directly influences CNN classification accuracy. While Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e indeed illustrates a range of bark morphologies among our selected species, from highly distinct to subtly similar patterns, this was a deliberate and integral aspect of our experimental design, not a fragility. This strategic selection allowed us to rigorously evaluate the evaluate the model's performance across a spectrum of bark morphological distinctiveness and to gain valuable \u003cem\u003einsights\u003c/em\u003e into the specific challenges CNN\u0026rsquo;s face when differentiating species based on bark. As hypothesized, \u003cem\u003eV. haenkeana\u003c/em\u003e, characterized by its more distinct and uniform textural patterns (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), achieved superior performance, including a precision rate of 100% and a recall of 94%. This high performance is likely attributed to \u003cem\u003eV. haenkeana\u003c/em\u003e smoother bark, generally lacking marked textures like prominent lenticels or fissures (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA-D). Conversely, \u003cem\u003eA. fraxinifolium\u003c/em\u003e and \u003cem\u003eA. leiocarpa\u003c/em\u003e exhibited higher misclassification rates due to their significant morphological similarity, particularly in shared features such as lenticels and scars, which present a more intricate and often convergent appearance. Specifically, \u003cem\u003eA. leiocarpa\u003c/em\u003e bark is typically fissured, rugose, and exfoliates rounded irregular plates, varying in color from brown to gray, whitish, yellowish, or reddish (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eE-H). Similarly, \u003cem\u003eA. fraxinifolium\u003c/em\u003e bark is fissured, rugose, and exfoliating in rounded irregular plates, with a color range from brown to gray, whitish, yellowish, or reddish (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eI-L).\u003c/p\u003e \u003cp\u003eThese results underscore the critical role of textural morphology in species differentiation by CNN\u0026rsquo;s, consistent with previous studies (Warner et al. \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; Warner et al. \u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e2024b\u003c/span\u003e). While bark characteristics are known to vary with ontogeny (e.g., thickness and rugosity changing with tree age/size), our sampling approach, which focused on mature individuals with a minimum DBH of 25 cm, aimed to standardize the developmental stage of the sampled trees. However, residual intra-specific variation related to ontogeny could still influence classification performance and represent an inherent complexity of natural datasets.\u003c/p\u003e \u003cp\u003eOur results further confirm the significant influence of common surface elements and environmental variability on the model's classification performance, directly addressing our second research question. The observed greater difficulty in classifying \u003cem\u003eA. fraxinifolium\u003c/em\u003e and \u003cem\u003eA. leiocarpa\u003c/em\u003e is notably linked to their shared superficial characteristics, such as lenticels, protuberances, and scars (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eC). This challenge is exacerbated by the seasonal abundance of epiphytes like lichens and bryophytes during the rainy season, which introduce substantial visual noise and confounding features (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eC). These elements can obscure subtle bark patterns, effectively blurring the visual distinctions between species, thus hindering CNN's ability to extract precise discriminatory features. This highlights a critical challenge for automated identification systems operating in dynamic natural environments, where visual heterogeneity is inherent and fluctuates with environmental conditions.\u003c/p\u003e \u003cp\u003eIn comparison to previous studies, this model demonstrates notable performance, achieving 91% overall accuracy in a highly biodiverse ecosystem like the Cerrado, which surpasses the average accuracy of 89% reported in similar CNN-based tree species identification efforts (Liu et al. \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). This high performance underscores the practical applicability of our model in supporting large-scale forest inventories and environmental monitoring initiatives. Specifically, the model's capacity to rapidly and accurately assimilate species with simpler textural patterns, as observed for \u003cem\u003eV. haenkeana\u003c/em\u003e, suggests its immediate utility for initial species screening in extensive projects. Indeed, our findings corroborate the significance potential of AI-based tools to enhance accuracy in forest inventories and support decision-making in tropical forest management (Gama et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eDespite the promising performance, this study acknowledges inherent limitations that offer avenues for future research. Firstly, while focusing on mature trees with a standardized DBH minimized ontogenetic variation, comprehensive assessment of temporal patterns and a broader representation of environmental variability (e.g., light conditions, canopy openness across seasons beyond just rainy season) were not exhaustively captured in the current dataset. Secondly, the model's reliance on bark images alone, while advantageous for phenology-independent identification, means it does not integrate other potentially diagnostic features (e.g., leaf, flower, fruit morphology, or phylogenetic context), which could further enhance accuracy, particularly for species with highly convergent bark patterns.\u003c/p\u003e \u003cp\u003eTo overcome these limitations and further advance the application of AI in ecological research, future work should prioritize expanding dataset size and diversity to include multiple seasons and broader geographical ranges. Additionally, integrating human expertise for iterative model refinement is crucial. Exploring multimodal deep learning approaches that combine various plant features (e.g., bark, leaf, flower, fruit morphology, or phylogenetic context) could significantly enhance accuracy, applicability, and robustness in diverse and dynamic ecosystems such as the Cerrado(Lapkovskis et al. \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Yang et al. \u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e2025b\u003c/span\u003e). This is a pertinent investigation, as traditional dichotomous keys for family and genus identification are primarily based on reproductive organs, with limited indications that bark patterns (rhytidoma) could be reliably used. However, the aid of AI, particularly in macro analysis, presents a promising avenue to investigate if structural rhytidome patterns and textural patterns are linked to families and genera, thereby expanding the utility of bark-based identification.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study demonstrates the strong potential of the MobileNetV2 model for tree species identification in the Brazilian Cerrado using bark images, achieving an overall accuracy of 90.52%. The results confirm that bark morphological complexity plays a key role in classification performance: species with distinctive bark patterns, such as \u003cem\u003eVochysia haenkeana\u003c/em\u003e, were classified more accurately, whereas species with convergent characteristics, including \u003cem\u003eAstronium fraxinifolium\u003c/em\u003e and \u003cem\u003eApuleia leiocarpa\u003c/em\u003e, posed greater challenges. The presence of superficial elements (e.g., lichens and scars) and environmental variability also affected model performance by introducing visual noise. Despite these challenges, MobileNetV2 represents a valuable complementary tool for large-scale forest inventories and environmental monitoring. Its phenology-independent approach is particularly advantageous in seasonally dynamic biomes like the Cerrado, especially where botanical expertise is limited, contributing to more efficient biodiversity assessments and conservation planning. However, this study has limitations, notably the use of data from a single seasonal period and the focus on species-level classification, highlighting the need for broader temporal coverage and taxonomic scope. Future research should expand the dataset across seasons and regions, integrate expert knowledge for iterative model refinement, and explore multimodal deep learning approaches combining multiple plant traits. Such advances could enhance accuracy, robustness, and the potential to extend bark-based identification beyond the species level, reinforcing the role of AI in biodiversity conservation.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJS Reis would like to thank CAPES (Coordenação de Aperfeiçoamento de Pessoal de Nível Superior) for the scholarship. CM Silva-Neto thanks CNPq (Conselho Nacional de Desenvolvimento Científico e Tecnológico) for the productivity grant. E Tizo-Pedroso thanks CNPq (408977/2016-7), Fundação de Amparo à Pesquisa do Estado de Goiás (FAPEG) (201610267001020), and Universidade Estadual de Goiás (Programa Pró-Pesquisa, 202200020022764) (Projetos Estratégicos, 202400020016002) for funding grants.\u003c/p\u003e\n\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eFunding was provided to E. Tizo-Pedroso by Universidade Estadual de Goiás (Programa Pró-Pesquisa, 202200020022764) (Projetos Estratégicos, 202400020016002), by Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq, 408977/2016-7), and ), Fundação de Amparo à Pesquisa do Estado de Goiás (FAPEG) (201610267001020).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCRediT authorship contribution statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJ.S.R: conceptualization, methodology, visualization, validation, data curation, formal analysis, writing-original draft. M.H.A.D.: Data curation and Formal analysis. C.M.S.N. and E.T.P.: supervision, resources, validation, writing-review and editing. All authors reviewed the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of Interest Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets generated during and/or analyzed during the conduction of the study are available from the corresponding author upon reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAggarwal CC (2018) Convolutional Neural Networks. In: Aggarwal CC (ed) Neural Networks and Deep Learning: A Textbook. Springer International Publishing, Cham, pp 315\u0026ndash;371. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/978-3-319-94463-0_8\u003c/span\u003e\u003cspan address=\"10.1007/978-3-319-94463-0_8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAkay M et al (2021) Deep Learning Classification of Systemic Sclerosis Skin Using the MobileNetV2 Model. IEEE Open J Eng Med Biology 2:104\u0026ndash;110. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/OJEMB.2021.3066097\u003c/span\u003e\u003cspan address=\"10.1109/OJEMB.2021.3066097\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAl-Gaashani MSAM, Xu W, Obsie EY (2025) MobileNetV2-based deep learning architecture with progressive transfer learning for accurate monkeypox detection. Appl Soft Comput 169:112553. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.asoc.2024.112553\u003c/span\u003e\u003cspan address=\"10.1016/j.asoc.2024.112553\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAltal OF et al (2025) Hybrid attention-enhanced MobileNetV2 with particle swarm optimization for endometrial cancer classification in CT images Informatics. Med Unlocked 57:101662. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.imu.2025.101662\u003c/span\u003e\u003cspan address=\"10.1016/j.imu.2025.101662\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnselin L, Syabri I, Kho Y (2005) Anal Geographical Anal 38:5\u0026ndash;22. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/j.0016-7363.2005.00671.x\u003c/span\u003e\u003cspan address=\"10.1111/j.0016-7363.2005.00671.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. GeoDa: An Introduction to Spatial Data\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eArora V, Rawat D, Singh N, Dwivedi SP, Chaturvedi P (2025) Leaf Based Plant Identification Using Traditional and Automated Methods. Rev Plant Archives 25:495\u0026ndash;509. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.51470/PLANTARCHIVES.2025.v25.supplement-1.067\u003c/span\u003e\u003cspan address=\"10.51470/PLANTARCHIVES.2025.v25.supplement-1.067\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBaihaqi MN, Sugiharto A, Tantyoko H (2025) Classification of Real and Fake Images Using Error Level Analysis Technique and MobileNetV2 Architecture. Jurnal Masyarakat Informatika 16:54\u0026ndash;68. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.14710/jmasif.16.1.73283\u003c/span\u003e\u003cspan address=\"10.14710/jmasif.16.1.73283\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBerte ML, Quirino CB, Fran\u0026ccedil;oso R (2024) The economic potential of non-timber forest products in the Cerrado legal reserves - El potencial econ\u0026ocirc;mico de los productos forestales no madereros en las reservas legales de. El Cerrado Caldasia 46:381\u0026ndash;394. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.15446/caldasia.v46n2.102712\u003c/span\u003e\u003cspan address=\"10.15446/caldasia.v46n2.102712\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBhusnurmath RA, Doddamani S Bark Texture Classification Using Deep Transfer Learning. In: Morusupalli R, Dandibhotla TS, Atluri VV, Windridge D, Lingras P, Komati VR (eds) Multi-disciplinary Trends in Artificial Intelligence, Cham, 2023// 2023. Springer Nature Switzerland, pp 407\u0026ndash;420. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/978-3-031-36402-0_38\u003c/span\u003e\u003cspan address=\"10.1007/978-3-031-36402-0_38\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBlaanco LJ, Travieso CM, Quinteiro JM, Hernandez PV, Dutta MK (2016) Singh A A bark recognition algorithm for plant classification using a least square support vector machine. In: Ninth International Conference on Contemporary Computing (IC3), 11\u0026ndash;13 Aug. 2016 2016. pp 1\u0026ndash;5. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/IC3.2016.7880233\u003c/span\u003e\u003cspan address=\"10.1109/IC3.2016.7880233\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCaccianiga M, Compostella C, Caccia G, Cattaneo C (2021) Contribution of plant anatomy to forensic investigation: Tree bark morphology. Forensic Sci Int 318:110598. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.forsciint.2020.110598\u003c/span\u003e\u003cspan address=\"10.1016/j.forsciint.2020.110598\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCicero M, Bilbily A, Colak E, Dowdell T, Gray B, Perampaladas K, Barfett J (2017) Training and Validating a Deep Convolutional Neural Network for Computer-Aided Detection and Classification of Abnormalities on Frontal Chest. Radiographs Invest Radiol 52. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/RLI.0000000000000341\u003c/span\u003e\u003cspan address=\"10.1097/RLI.0000000000000341\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eColli GR, Vieira CR, Dianese JC (2020) Biodiversity and conservation of the Cerrado: recent advances and old challenges. Biodivers Conserv 29:1465\u0026ndash;1475. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s10531-020-01967-x\u003c/span\u003e\u003cspan address=\"10.1007/s10531-020-01967-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElgamily KM, Mohamed MA, Abou-Taleb AM, Ata MM (2025) A novel W13 deep CNN structure for improved semantic segmentation of multiple objects in remote sensing imagery Neural. Comput Appl 37:5397\u0026ndash;5427. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s00521-024-10765-3\u003c/span\u003e\u003cspan address=\"10.1007/s00521-024-10765-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEnr\u0026iacute;quez-de-Salamanca \u0026Aacute; Botanical databases in EIA: opportunities and challenges Impact Assessment and Project Appraisal:1\u0026ndash;11 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/14615517.2025.2482137\u003c/span\u003e\u003cspan address=\"10.1080/14615517.2025.2482137\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFan C-L (2025) Evaluation Model for Crack Detection with Deep Learning: Improved Confusion Matrix Based on Linear Features. J Constr Eng Manag 151:04024210. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1061/JCEMD4.COENG-14976\u003c/span\u003e\u003cspan address=\"10.1061/JCEMD4.COENG-14976\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFekri-Ershad S (2020) Bark texture classification using improved local ternary patterns and multilayer neural network Expert Systems with Applications 158:113509 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.eswa.2020.113509\u003c/span\u003e\u003cspan address=\"10.1016/j.eswa.2020.113509\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGama NC et al (2025) Automatic Recognition of Commercial Tree Species from the Amazon Flora Using Bark Images and Transfer Learning. Preprints. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.20944/preprints202506.1393.v1\u003c/span\u003e\u003cspan address=\"10.20944/preprints202506.1393.v1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGulzar Y (2023) Fruit Image Classification Model Based on MobileNetV2 with Deep Transfer Learning Technique. Sustainability 15:1906. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/su15031906\u003c/span\u003e\u003cspan address=\"10.3390/su15031906\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHajjaji Y, Boulila W, Farah IR, Koubaa A (2025) Enhancing palm precision agriculture: An approach based on deep learning and UAVs for efficient palm tree detection. Ecol Inf 85:102952. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ecoinf.2024.102952\u003c/span\u003e\u003cspan address=\"10.1016/j.ecoinf.2024.102952\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHortal J, De-Bello F, Diniz-Filho JAF, Lewinsohn TM, Lobo JM, Ladle RJ (2015) Seven shortfalls that beset large-scale knowledge of biodiversity Annual review of ecology. Evol Syst 46:523\u0026ndash;549. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1146/annurev-ecolsys-112414-054400\u003c/span\u003e\u003cspan address=\"10.1146/annurev-ecolsys-112414-054400\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuo R, Glickman ME (2025) Receiver operating characteristic analysis for paired comparison data. J Royal Stat Soc Ser A: Stat Soc 188:741\u0026ndash;764. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/jrsssa/qnae072\u003c/span\u003e\u003cspan address=\"10.1093/jrsssa/qnae072\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIndraswari R, Rokhana R, Herulambang W (2022) Melanoma image classification based on MobileNetV2 network. Procedia Comput Sci 197:198\u0026ndash;207. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.procs.2021.12.132\u003c/span\u003e\u003cspan address=\"10.1016/j.procs.2021.12.132\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKatsis LKD, Hill AP, Pi\u0026ntilde;a-Covarrubias E, Prince P, Rogers A, Patrick Doncaster C, Snaddon JL (2022) Automated detection of gunshots in tropical forests using convolutional neural networks. Ecol Ind 141:109128. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ecolind.2022.109128\u003c/span\u003e\u003cspan address=\"10.1016/j.ecolind.2022.109128\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKattenborn T, Leitloff J, Schiefer F, Hinz S (2021) Review on Convolutional Neural Networks (CNN) in vegetation remote sensing ISPRS. J Photogrammetry Remote Sens 173:24\u0026ndash;49. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.isprsjprs.2020.12.010\u003c/span\u003e\u003cspan address=\"10.1016/j.isprsjprs.2020.12.010\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim TK et al (2022) Identifying and extracting bark key features of 42 tree species using convolutional neural networks and class activation mapping Scientific reports 12:4772 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-022-08571-9\u003c/span\u003e\u003cspan address=\"10.1038/s41598-022-08571-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumar T, Brennan R, Mileo A, Bendechache M (2024) Image Data Augmentation Approaches: A Comprehensive Survey and Future. Dir IEEE Access 4:1\u0026ndash;1. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ACCESS.2024.3470122\u003c/span\u003e\u003cspan address=\"10.1109/ACCESS.2024.3470122\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumaresan S, Aultrin KSJ, Kumar SS, Anand MD (2021) Transfer Learning With CNN for Classification of Weld. Defect IEEE Access 9:95097\u0026ndash;95108. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ACCESS.2021.3093487\u003c/span\u003e\u003cspan address=\"10.1109/ACCESS.2021.3093487\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLapkovskis A, Nefedova N, Beikmohammadi A (2024) Automatic Fused Multimodal Deep Learning for Plant Identification arXiv preprint arXiv. 240601455:1\u0026ndash;26. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48550/arXiv.2406.01455\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2406.01455\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeiva-Bianchi M, Castillo N, Astudillo CA, Ahumada-M\u0026eacute;ndez F (2025) Effectiveness of machine learning methods in detecting grooming: a systematic meta-analytic. Rev Sci Rep 15:9008. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-024-83003-4\u003c/span\u003e\u003cspan address=\"10.1038/s41598-024-83003-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi Z, Yang W, Peng S, Liu F (2020) A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.2004.02806\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2004.02806\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLira-Martins D et al (2022) Soil properties and geomorphic processes influence vegetation composition, structure, and function in the Cerrado. Domain Plant Soil 476:549\u0026ndash;588. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11104-022-05517-y\u003c/span\u003e\u003cspan address=\"10.1007/s11104-022-05517-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu J, Wang X, Wang T (2019) Classification of tree species and stock volume estimation in ground forest images using Deep. Learn Computers Electron Agric 166:105012. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.compag.2019.105012\u003c/span\u003e\u003cspan address=\"10.1016/j.compag.2019.105012\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu J, Xing T, Wang X (2025) Optimizing AlexNet for accurate tree species classification via multi-branch architecture and mixed-domain attention. Complex Intell Syst 11:244. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s40747-025-01836-6\u003c/span\u003e\u003cspan address=\"10.1007/s40747-025-01836-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMartello F et al (2023) Landscape structure shapes the diversity of plant reproductive traits in agricultural landscapes in the Brazilian Cerrado Agriculture. Ecosyst Environ 341:108216. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.agee.2022.108216\u003c/span\u003e\u003cspan address=\"10.1016/j.agee.2022.108216\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMeyer C, Weigelt P, Kreft H (2016) Multidimensional biases, gaps and uncertainties in global plant occurrence information. Ecol Lett 19:992\u0026ndash;1006. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/ele.12624\u003c/span\u003e\u003cspan address=\"10.1111/ele.12624\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMiftahushudur T, Sahin HM, Grieve B, Yin H (2025) A survey of methods for addressing imbalance data problems in. Agric Appl Remote Sens 17:454. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/rs17030454\u003c/span\u003e\u003cspan address=\"10.3390/rs17030454\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMisra D, Crispim-Junior C, Tougne L, Patch-Based CNN Evaluation for Bark Classification. In: Bartoli A, Fusiello A (eds) Computer Vision \u0026ndash; ECCV 2020 Workshops, Cham, 2020// 2020. Springer International Publishing, pp 197\u0026ndash;212. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/978-3-030-65414-6_15\u003c/span\u003e\u003cspan address=\"10.1007/978-3-030-65414-6_15\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMohammad AA, Dara R, Sunil K (2025) 7 - t-Distributed stochastic neighbor embedding. In: Amani Rad J, Chakraverty S, Parand K (eds) Dimensionality Reduction in Machine Learning. Morgan Kaufmann, pp 187\u0026ndash;207. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/B978-0-44-332818-3.00017-4\u003c/span\u003e\u003cspan address=\"10.1016/B978-0-44-332818-3.00017-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoran M, Faria M, Giraldi G, Bastos L, Oliveira L, Conci A (2021) Classification of Approximal Caries in Bitewing Radiographs Using Convolutional Neural Networks. Sensors 21. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/s21155192\u003c/span\u003e\u003cspan address=\"10.3390/s21155192\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMustin K et al (2017) Biodiversity, threats and conservation challenges in the Cerrado of Amap\u0026aacute;, an Amazonian. savanna Nat Conserv 22:107\u0026ndash;127. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3897/natureconservation.22.13823\u003c/span\u003e\u003cspan address=\"10.3897/natureconservation.22.13823\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNogueira K, Santos JA, Menini N, Silva TSF, Morellato LPC, Torres RS (2019) Spatio-Temporal Vegetation Pixel Classification by Using Convolutional Networks IEEE Geoscience and. Remote Sens Lett 16:1665\u0026ndash;1669. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/LGRS.2019.2903194\u003c/span\u003e\u003cspan address=\"10.1109/LGRS.2019.2903194\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNooralishahi P et al (2021) Drone-Based Non-Destructive Inspection of Industrial Sites: A Review and Case Studies. Drones 5. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/drones5040106\u003c/span\u003e\u003cspan address=\"10.3390/drones5040106\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQGIS (2022) Development Team QGIS Geographic Information System. Open Source Geospatial Foundation Project. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://qgis.osgeo.org\u003c/span\u003e\u003cspan address=\"http://qgis.osgeo.org\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Accessed 12 September 2022\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQuach BM, Dinh VC, Pham N, Huynh D, Nguyen BT (2023) Leaf recognition using convolutional neural networks based features Multimedia. Tools Appl 82:777\u0026ndash;801. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s11042-022-13199-y\u003c/span\u003e\u003cspan address=\"10.1007/s11042-022-13199-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRabeling SC, Lim JL, Tidon R, Neff JL, Simpson BB, Pawar S (2019) Seasonal variation of a plant-pollinator network in the Brazilian Cerrado: Implications for community structure and robustness. PLoS ONE 14:e0224997. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1371/journal.pone.0224997\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0224997\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRocha LGS, Pinto JRR (2021) Diversity and structural changes in Cerrado Rupestre under effects of disturbances Brazilian. J Bot 44:213\u0026ndash;225. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s40415-020-00687-4\u003c/span\u003e\u003cspan address=\"10.1007/s40415-020-00687-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRosell JA, Gleason S, M\u0026eacute;ndez-Alonzo R, Chang Y, Westoby M (2014) Bark functional ecology: evidence for tradeoffs, functional coordination, and environment producing bark diversity. New Phytol 201:486\u0026ndash;497. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/nph.12541\u003c/span\u003e\u003cspan address=\"10.1111/nph.12541\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRouhan G, Gaudeul M (2014) Plant Taxonomy: A Historical Perspective, Current Challenges, and Perspectives. In: Besse P (ed) Molecular Plant Taxonomy: Methods and Protocols. Humana, Totowa, NJ, pp 1\u0026ndash;37. doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/978-1-62703-767-9_1\u003c/span\u003e\u003cspan address=\"10.1007/978-1-62703-767-9_1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRzanny M, M\u0026auml;der P, Deggelmann A, Chen M, W\u0026auml;ldchen J (2019) Flowers, leaves or both? How to obtain suitable images for automated plant identification. Plant Methods 15:77. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s13007-019-0462-4\u003c/span\u003e\u003cspan address=\"10.1186/s13007-019-0462-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchwaida SF, Cicerelli RE, Almeida T, Sano EE, Pires CH, Ramos APM (2023) Defining priorities areas for biodiversity conservation and trading forest certificates in the Cerrado biome in. Brazil Biodivers Conserv 32:1807\u0026ndash;1820. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10531-023-02578-y\u003c/span\u003e\u003cspan address=\"10.1007/s10531-023-02578-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShah P, Mishra DK, Shanmugam MP, Doshi B, Jayaraj H, Ramanjulu R (2020) Validation of Deep Convolutional Neural Network-based algorithm for detection of diabetic retinopathy \u0026ndash; Artificial intelligence versus clinician for screening. Indian J Ophthalmol 68. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.4103/ijo.IJO_966_19\u003c/span\u003e\u003cspan address=\"10.4103/ijo.IJO_966_19\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShorten C, Khoshgoftaar TM, Furht B (2021) Text Data Augmentation for Deep Learning. J Big Data 8:101. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s40537-021-00492-0\u003c/span\u003e\u003cspan address=\"10.1186/s40537-021-00492-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSilva-Pereira BA, Venturoli F, Carvalho FA (2011) Florestas Estacionais no Cerrado: Uma Vis\u0026atilde;o. Geral Pesquisa Agropecu\u0026aacute;ria Trop 446\u0026ndash;455. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5216/pat.v41i3.12666\u003c/span\u003e\u003cspan address=\"10.5216/pat.v41i3.12666\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSilveira P et al (2025) The loss of an unknown biodiversity: Spatial gaps in plant survey and conservation in a Brazilian. hotspot Biodivers Biol Conserv 305:111098. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.biocon.2025.111098\u003c/span\u003e\u003cspan address=\"10.1016/j.biocon.2025.111098\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSouza KR, Oliveira G, Ferreira F, Martins TO, Martins T, Venturoli F, Silva-Neto C (2016a) Floresta Estacional: madeira para o produtor rural no Cerrado Seasonal Forest: Sustainable wood for farmers in. Brazilian Savanna Magistra 28:548\u0026ndash;556\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSouza KR, Oliveira GM, Ferreira FG, Martins TO, Melo C, Neto S, Venturoli F (2016b) Floresta Estacional: madeira sustent\u0026aacute;vel para o produtor. rural no Cerrado MAGISTRA 28:54\u0026ndash;62\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSurendran GK, Deekshitha, Lukac M, Lukac M, Vybostok J, Mokros M (2025) A forestry investigation: Exploring factors behind improved tree species classification using bark images. Ecol Inf 85:102932. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ecoinf.2024.102932\u003c/span\u003e\u003cspan address=\"10.1016/j.ecoinf.2024.102932\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang C et al (2022) Application of Convolutional Neural Network-Based Detection Methods in Fresh Fruit Production: A Comprehensive Review. Front Plant Sci 13. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fpls.2022.868745\u003c/span\u003e\u003cspan address=\"10.3389/fpls.2022.868745\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWarner C, Wu F, Gazo R, Benes B, Fei S (2025) Environmental Sensitivity in AI Tree Bark Detection: Identifying Key Factors for Improving Classification Accuracy. Algorithms 18. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/a18070417\u003c/span\u003e\u003cspan address=\"10.3390/a18070417\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWarner C, Wu F, Gazo R, Benes B, Kong N, Fei S (2024a) Central Bark Image Dataset and Tree Species Classification Using Deep Learning. Algorithms 17. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/a17050179\u003c/span\u003e\u003cspan address=\"10.3390/a17050179\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWarner C, Wu F, Gazo R, Benes B, Kong N, Fei S (2024b) CentralBark Image Dataset and Tree Species Classification Using Deep Learning. Algorithms 17. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/a17050179\u003c/span\u003e\u003cspan address=\"10.3390/a17050179\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu F, Gazo R, Benes B, Haviarova E (2021) Deep BarkID: a portable tree bark identification system by knowledge distillation European. J For Res 140:1391\u0026ndash;1399. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s10342-021-01407-7\u003c/span\u003e\u003cspan address=\"10.1007/s10342-021-01407-7\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu Y, Goodacre R (2018) On Splitting Training and Validation Set: A Comparative Study of Cross-Validation, Bootstrap and Systematic Sampling for Estimating the Generalization Performance of Supervised Learning. J Anal Test 2:249\u0026ndash;262. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s41664-018-0068-2\u003c/span\u003e\u003cspan address=\"10.1007/s41664-018-0068-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang G, Wang G, Wan L, Wang X, He Y (2025a) Utilizing SMOTE-TomekLink and machine learning to construct a predictive model for elderly medical and daily care services demand. Sci Rep 15:8446. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-025-92722-1\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-92722-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang Z-X, Li Y, Wang R-F, Hu P, Su W-H (2025b) Deep Learning in Multimodal Fusion for Sustainable Plant Care: A Comprehensive Review Sustainability (2071\u0026thinsp;\u0026ndash;\u0026thinsp;1050). 17. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/su17125255\u003c/span\u003e\u003cspan address=\"10.3390/su17125255\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y, Wang Z, Lu Y, Zuo L (2022) Biodiversity, ecosystem functions and services: Interrelationship with environmental and human health. Front Ecol Evol 10:1086408. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fevo.2022.1086408\u003c/span\u003e\u003cspan address=\"10.3389/fevo.2022.1086408\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Z, Ao Z, Wu W, Wang Y, Xin Q (2024) Developing a Multi-Scale Convolutional Neural Network for Spatiotemporal Fusion to Generate MODIS-like Data Using AVHRR and Landsat Images. Remote Sens 16. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/rs16061086\u003c/span\u003e\u003cspan address=\"10.3390/rs16061086\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Morphological Texture Analysis, Environmental Monitoring, Image Classification Models, Tropical Ecosystems","lastPublishedDoi":"10.21203/rs.3.rs-8841518/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8841518/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe Brazilian Cerrado, a global biodiversity hotspot, faces persistent challenges in species identification due to the limitations of phenology‑dependent methods in its seasonally dynamic landscapes. This study evaluates the performance of convolutional neural networks (CNNs) for bark‑image‑based classification of three ecologically prominent tree species: \u003cem\u003eApuleia leiocarpa\u003c/em\u003e, \u003cem\u003eAstronium fraxinifolium\u003c/em\u003e, and \u003cem\u003eVochysia haenkeana\u003c/em\u003e. We compiled 1,515 bark images from individual trees (DBH\u0026thinsp;\u0026ge;\u0026thinsp;25 cm) during the 2023\u0026ndash;2024 rainy seasons and applied data augmentation and normalization. Using the MobileNetV2 architecture, we trained and validated the model with metrics including Accuracy, Precision, Recall, F1‑score, Confusion Matrix, ROC/AUC curves, and t‑SNE projections. The model achieved an overall accuracy of 90.52%. Bark morphological complexity strongly influenced classification: \u003cem\u003eV. haenkeana\u003c/em\u003e, with distinct patterns, showed the highest performance (Precision 1.00, Recall 0.94), while \u003cem\u003eA. fraxinifolium\u003c/em\u003e and \u003cem\u003eA. leiocarpa\u003c/em\u003e, which share more convergent bark traits, exhibited higher misclassification rates (22.54%). These results demonstrate how interspecific bark variability affects CNN discrimination and confirm that intrinsic bark heterogeneity (e.g., rhytidome texture, rugosity, color patterns, scars) and environmental variation increase classification difficulty. Our findings highlight the potential of bark‑based deep learning models as phenology‑independent tools for large‑scale forest inventories and biodiversity monitoring in complex ecosystems. A key limitation is the dataset\u0026rsquo;s restriction to a single seasonal period, underscoring the need for broader temporal sampling. This study reinforces the role of deep learning in delivering scalable and accurate solutions for ecological research and conservation in understudied biodiversity hotspots.\u003c/p\u003e","manuscriptTitle":"Overcoming Structural and Environmental Challenges in Identifying Brazilian Savanna Tree Species Using Deep Learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-18 17:04:10","doi":"10.21203/rs.3.rs-8841518/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"ea3fa1ba-e033-4da9-9344-016ea279960d","owner":[],"postedDate":"February 18th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-02-18T17:04:10+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-18 17:04:10","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8841518","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8841518","identity":"rs-8841518","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}