Expert Knowledge-Guided Decision Calibration for Accurate Fine-Grained Tree Species Classification

TL;DR

This work tackles the challenge of fine-grained tree species classification under long-tailed distributions and high inter-class similarity by introducing EKDC-Net, which leverages an external domain expert to recalibrate backbone decisions. The method comprises Local Prior-guided Knowledge Extraction (LPKEM) to focus expert analysis on discriminative regions via CAM cues and Uncertainty-guided Decision Calibration (UDCM) to dynamically weight expert input using both class-level and instance-level uncertainty; the calibrated logit is given by $\hat{\boldsymbol{z}} = \boldsymbol{z}_b + \lambda \cdot \boldsymbol{z}_e$. A frozen BioCLIP-based expert provides robust global priors, while LPKEM filters background noise and selectively exposes relevant features. The authors introduce CU-Tree102 (9,134 images, 102 species) and demonstrate state-of-the-art results on CU-Tree102, RSTree, and Jekyll with a lightweight overhead of only $0.08$M parameters, achieving notable gains in accuracy and macro metrics, particularly for rare classes. The work offers a practical, generalizable approach to forest inventory and urban forestry tasks by effectively combining local discriminative power with broad expert knowledge.

Abstract

Accurate fine-grained tree species classification is critical for forest inventory and biodiversity monitoring. Existing methods predominantly focus on designing complex architectures to fit local data distributions. However, they often overlook the long-tailed distributions and high inter-class similarity inherent in limited data, thereby struggling to distinguish between few-shot or confusing categories. In the process of knowledge dissemination in the human world, individuals will actively seek expert assistance to transcend the limitations of local thinking. Inspired by this, we introduce an external "Domain Expert" and propose an Expert Knowledge-Guided Classification Decision Calibration Network (EKDC-Net) to overcome these challenges. Our framework addresses two core issues: expert knowledge extraction and utilization. Specifically, we first develop a Local Prior Guided Knowledge Extraction Module (LPKEM). By leveraging Class Activation Map (CAM) analysis, LPKEM guides the domain expert to focus exclusively on discriminative features essential for classification. Subsequently, to effectively integrate this knowledge, we design an Uncertainty-Guided Decision Calibration Module (UDCM). This module dynamically corrects the local model's decisions by considering both overall category uncertainty and instance-level prediction uncertainty. Furthermore, we present a large-scale classification dataset covering 102 tree species, named CU-Tree102 to address the issue of scarce diversity in current benchmarks. Experiments on three benchmark datasets demonstrate that our approach achieves state-of-the-art performance. Crucially, as a lightweight plug-and-play module, EKDC-Net improves backbone accuracy by 6.42% and precision by 11.46% using only 0.08M additional learnable parameters. The dataset, code, and pre-trained models are available at https://github.com/WHU-USI3DV/TreeCLS.

Figures (10)

• Figure 1: Illustration of the limitations in existing methods. (a) Struggle with Long-tail Categories: The upper histogram illustrates the imbalanced sample distribution, while the lower plot reveals the corresponding drop in classification accuracy. (b) Ambiguity in Local Distributions: Although the global distribution exhibits inherent separability, the local distribution reveals significant overlap and ambiguity along the decision boundary, due to the limitations of the data.
• Figure 2: Overview of EKDC-Net architecture. Given a tree image, EKDC-Net accurately identifies the fine-grained tree species.
• Figure 3: An illustration of the Local Prior-guided Knowledge Extraction Module. means frozen this network parameters, denotes the trainable network parameters.
• Figure 4: An illustration of the Uncertainty-guided Decision Calibration Module.
• Figure 5: Accuracy comparison grouped by sample counts. We evaluate the proposed EKDC-Net against three state-of-the-art fine-grained methods: (a) MPSA (b) HERBS, and (c) CGL; and three standard backbones: (d) ViT-Base, (e) Swin-Base, and (f) ResNet-50. The proposed method (orange) consistently outperforms the baseline (blue). Percentages represent relative improvement, and * indicates the intervals with the greatest increase.