Explainable deep learning classification of tree species with very high resolution VHRTreeSpecies dataset

Abstract

Forests are among the most vital natural resources, playing a significant role in regulating the climate, maintaining ecological balance, and supporting biodiversity conservation and sustainable forest management. Additionally, they contribute to various applications, such as hazard management and wildlife habitat mapping. Understanding the spatial and temporal distribution of forests and forest stand types is a prerequisite for gaining deeper insights into their role within the Earth's systems. In this context, remote sensing data is widely utilized for forest stand type classification. However, traditional classification methods are often time-consuming and typically limited to specific areas and species, which significantly restricts their applicability to different regions and diverse tree species. With the increasing availability of high-resolution satellite imagery, deep learning methods have emerged as a powerful tool for forest management and tree species classification, offering enhanced efficiency and broader applicability compared to conventional approaches. Remote sensing (RS) applications, which serve as an essential spatial data source in forestry practices, have emerged as an effective solution for field studies due to their cost-efficiency and rapid data acquisition capabilities. Remote sensing systems provide valuable spatial, temporal, and spectral resolution data to cover forest areas at the required scale and within the necessary temporal intervals for data collection. High-resolution remote sensing data are frequently preferred for deriving detailed tree-level information, particularly for tasks such as individual tree detection or damage assessment necessary for maintaining tree health. Satellite systems such as Sentinel-2 and Landsat are frequently preferred due to their open-access nature, which allows for the collection of data across broad spectral bands and the provision of continuous data access. Nevertheless, the spatial resolution limitations of these satellites may render them inadequate for particular applications. Tree species with varying structural and morphological characteristics exhibit distinct spectral properties. Trees within the same environment but at varying developmental stages or health conditions can display significant differences in their spectral characteristics. In this regard, the application of remote sensing data is essential for achieving precise and reliable classification of tree species. Over the past decade, considerable progress has been made in the identification of tree species, encompassing a spectrum of approaches from fundamental image processing techniques to sophisticated machine learning (ML) and deep learning (DL) methodologies. Nevertheless, traditional classification algorithms, such as Random Forest (RF) and Support Vector Machines (SVM), have shown limited effectiveness in identifying tree canopies within dense and complex backgrounds. However, the time-consuming nature of traditional methods and their typical application to only specific areas and tree species substantially constrain the usability of these models across different regions and diverse species. Conversely, with the increasing availability of high-resolution satellite imagery, deep learning methods have emerged as powerful tools in forest management and tree species classification. DL-based models possess the potential to accurately extract more intricate information structures. Nevertheless, the effective application of these models generally requires a larger number of reference data samples to enable sufficient learning of the model parameters. As part of this thesis, a new benchmark dataset for forest stand type classification, called VHRTreeSpecies, is introduced. This comprehensive dataset includes very high-resolution RGB satellite imagery of 15 dominant tree species from various forest ecosystems across Turkey. The input images and their corresponding labels were generated using Google Earth imagery and forest stand maps provided by the General Directorate of Forestry (GDF). The dataset was curated by selecting pure species and masking raster images using vector data. High-quality images captured during the summer months (late July to mid-August) from the past five years were prioritized. The dataset was further diversified to represent different forest stand development stages (youth, sapling, thin, medium, and mature trees) and canopy closure levels (open, moderately closed, fully closed). The dataset was analyzed using various CNN architectures, including ResNet-50, ResNet-101, VGG16, VGG19, ResNeXt-50, EfficientNet, and ConvNeXt. Additionally, explainable artificial intelligence (XAI) methods, such as Occlusion, Integrated Gradients and Grad-CAM, were applied to examine the decision-making processes of the models. Evaluation metrics, including Max-Sensitivity and AUC-MoRF, were employed to comprehensively assess model performance not only in terms of classification accuracy but also in terms of the interpretability and reliability of their decision-making mechanisms.