UruDendro4: A Benchmark Dataset for Automatic Tree-Ring Detection in Cross-Section Images of Pinus taeda L

AI summary

Overview

• Research area: Computer vision applied to forestry and dendrochronology — specifically automatic detection of annual tree-ring boundaries in wood cross-section photographs.
• Technical level: Intermediate. It requires familiarity with segmentation metrics (mAP, mAR, Adapted Rand error) and U-Net-based detectors, but the dataset description and acquisition protocol are accessible to non-specialists.
• Scope: This paper releases a 102-image annotated benchmark dataset of Pinus taeda L. cross-sections and benchmarks three state-of-the-art ring-detection methods on it, with a series of ablation experiments.

What This Paper Is About

Measuring annual tree growth usually means manually tracing rings along only 4 to 8 radial directions on a cross-section disc, which is slow and imprecise. The authors introduce UruDendro4, a public dataset of 102 manually annotated Pinus taeda L. cross-section images, to support the development and fair comparison of automatic ring-detection algorithms. A distinguishing feature is that it contains multiple cross-sections taken at different heights along the same stem, which makes it possible to model annual volume growth rather than just one-dimensional ring widths.

Key Contributions

1. A new public dataset. UruDendro4 contains 102 annotated cross-section images of Pinus taeda L. with 1930 annotated tree rings — the largest public annotated cross-section dataset reported in the paper's comparison table, in terms of both number of samples and number of rings.
2. Multi-height sampling for volumetric analysis. Unlike prior collections, the dataset includes discs extracted at heights of 0 m, 3 m, 6 m, and 9 m (plus an extra disc at 1.3 m for the T0 treatment), enabling volumetric modeling of annual growth from cross-sectional imagery.
3. Full documentation. The paper describes the image acquisition protocol (iPhone 15 Pro, outdoor and indoor lighting, moistened polished discs), the image naming scheme, the annotation methodology (Labelme, with initial predictions from CS-TRD and the APD pith detector, then human correction), and the silvicultural and meteorological metadata.
4. A performance baseline. Recent ring-boundary detection algorithms are benchmarked on the dataset, plus ablation experiments on training-set composition, encoder freezing, and parameter tuning, and a cross-species generalization test.

Main Findings

• DeepCS-TRD is the strongest baseline. On the UruDendro4 test set, Table III reports CS-TRD at .576 mAP / .568 mAR / .174 ARAND, INBD at .736 / .712 / .099, and DeepCS-TRD at .831 (.007) / .775 (.007) / .087 (.004). (The body text quotes the DeepCS-TRD row slightly differently — 0.829 ± 0.006 mAP, 0.769 ± 0.004 mAR, 0.094 ± 0.007 ARAND — so the paper is internally inconsistent on this row.)
• Tuning the edge-filter parameter helps. A grid search over the DeepCS-TRD α parameter (30, 45, 60, 75, 90 degrees) found the best result at α = 60°, reaching 0.838 ± 0.016 mAP, 0.782 ± 0.011 mAR, and 0.084 ± 0.006 ARAND. This α = 60° configuration is the headline result quoted in the abstract and conclusions.
• Freezing the encoder improves results. Training DeepCS-TRD on the UruDendro4 training set with frozen encoder weights gave .831 / .775 / .087 versus .829 / .769 / .094 with all layers trained, and reduced trainable parameters from 14,328,209 to 3,151,697.
• Pre-training on the earlier datasets transfers. Training only on UruDendro (Uru) gave .710 / .683 / .115, only on UruDendro2 (Uru2) gave .694 / .680 / .112, and the union (Uru + Uru2) gave .729 / .705 / .107 — all better than the training-free CS-TRD.
• Stacking datasets is not always better than the largest single set. Training on Uru + Uru2 + Uru4 gave .820 / .767 / .087 (frozen encoder: .806 / .750 / .102), below the UruDendro4-only frozen-encoder result. A two-step recipe (train on Uru + Uru2 for 100 epochs, then fine-tune on UruDendro4 for 100 epochs) gave .826 / .775 / .090, or .826 / .774 / .087 with the encoder frozen.
• Cross-species generalization works. Evaluated on two manually annotated Douglas fir images (79 tree rings), the model trained only on UruDendro4 scored .465 / .437 / .339, on Uru scored .629 / .623 / .208, on Uru2 scored .725 / .688 / .176, and on the combination Uru + Uru2 + Uru4 scored best at .728 (.013) / .705 (.021) / .145 (.009).
• Adding UruDendro4 to training improves generalization. The conclusions state that on the Douglas fir dataset, including UruDendro4 yielded gains of 0.003 in mAP, 0.017 in mAR, and 0.031 in ARAND.
• INBD is competitive but far more expensive to train. Despite being designed for microscopy images of shrubs, INBD reached .736 / .712 / .099 on UruDendro4, but required 18 hours of training versus 33 minutes for DeepCS-TRD.
• Per-sample qualitative results. For sample T6_B3_N21_D (which contains two knots), CS-TRD scored 0.256 mAP / 0.241 mAR / 0.487 ARAND, INBD 0.306 / 0.288 / 0.302, and DeepCS-TRD 0.688 / 0.688 / 0.245. For sample T0_B1_N32_ADAP, CS-TRD scored 0.588 / 0.470 / 0.200, INBD 0.677 / 0.610 / 0.135, and DeepCS-TRD 0.800 / 0.680 / 0.118.

Methodology in Plain English

The authors collected 102 cross-section discs from Pinus taeda L. trees at a site called "La Altura," roughly 4 km from Tranqueras in the Rivera department of Uruguay. The trees came from a randomized complete block design (blocks B1 to B3) testing four silvicultural treatments (T0, T2, T4, T6) that differ in thinning timing, post-thinning density, and pruning height; all treatments started from a pre-thinning density of 600 trees per hectare. Discs were photographed with an iPhone 15 Pro under a mix of indoor artificial light, outdoor light, and outdoor shade; image widths range from 1317 to 4695 pixels.

Annotations were produced by first running CS-TRD (combined with the APD pith detector) to generate initial ring curves, then having a human annotator correct them in the Labelme tool. Each ring is a continuous curve of connected pixels; rings are roughly concentric around the pith and do not intersect.

For the benchmark, images were preprocessed following the protocol from the earlier UruDendro2 work: background removed with a U²-Net network, margins cropped so the disc sits at least 50 pixels from the border, and images resized to 1504 × 1504 pixels on the longest dimension using Lanczos interpolation. The dataset was split 60%/20%/20% into 62 training, 20 validation, and 20 test images, and all reported numbers come from the test subset.

Three detectors were compared: CS-TRD (classical image processing, no training, σ = 3.0 and α = 45), DeepCS-TRD (a U-Net variant with a ResNet18 ImageNet backbone, Dice loss, Adam, batch size 8, 100 epochs, initial learning rate 1×10⁻³ with cosine annealing), and INBD (a two-stage U-Net approach with a MobileNetV3 ImageNet backbone, downsampling factor 2, AdamW, batch size 8, 100 epochs, initial learning rate 1×10⁻³ with cosine annealing, only decoder and classifier layers trainable). All models were trained on a single NVIDIA A40 GPU with 48 GB of VRAM via the ClusterUY infrastructure. Evaluation used mAP and mAR averaged over IoU thresholds from 0.5 to 1.0 in steps of 0.05, plus the Adapted Rand error computed with skimage's adapted_rand_error (version 0.25), reported as mean and standard deviation over five independent experiments.

Why This Matters

• Impact on research. Public annotated datasets are a bottleneck in this field. UruDendro4 nearly doubles the number of annotated P. taeda samples available, from 117 (64 in UruDendro plus 53 in UruDendro2) to 219, and its 1930 rings exceed the next-largest entries in the paper's comparison table. It gives researchers a larger, more diverse benchmark and a documented baseline to build on.
• Real-world applications:
  • Forestry companies evaluating whether specific thinning schedules and pruning regimes actually accelerate growth under local soil and climate conditions.
  • Dendrochronology and climate science, where ring-derived metrics are correlated against temperature and precipitation records (the dataset ships with INUMET meteorological data from January 2002 to April 2024).
  • Wood-quality and sawmill applications, where knowing ring geometry and volume affects processing decisions.
  • National forest inventory and carbon accounting, where per-tree volume increment estimates matter.
• Industry relevance. The paper notes the site has a productivity index of 53 and soils classified as priority forestry soils (Umbric Albic Luvisols, CONEAT group 7.42), which ties the work directly to commercial plantation management in Uruguay. Automated, two-dimensional and volumetric ring measurement could replace a process that is currently manual and limited to 4–8 radial measurement lines per disc, reducing the manual post-processing still needed to fix missing or false ring detections.

Future Directions

• Extending volumetric modeling. The multi-height sampling (0 m, 3 m, 6 m, 9 m, plus 1.3 m for T0) is explicitly framed as enabling volumetric annual-growth models, but the paper reports no volumetric result; building and validating such models is the obvious next step.
• Understanding why UruDendro4-only training can beat larger combined training sets. UruDendro4 alone with a frozen encoder (.831 / .775 / .087) outperformed the combined Uru + Uru2 + Uru4 setting, which the paper attributes to the pre-trained encoder's feature extraction capability. Characterizing when more data helps and when it hurts remains open.
• Broadening species coverage. Cross-domain tests currently rest on only two annotated Douglas fir images (79 rings), and the paper's comparison table lists datasets for Abies alba, Gleditsia triacanthos, Vaccinium myrtillus, Dryas octopetala, Empetrum hermaphroditum, and Salix glauca. Expanding annotated multi-species benchmarks would test generalization more rigorously.
• Reducing dependence on manual correction. Initial ring curves still require a human post-processing step to add missing rings and remove false detections; closing that gap, and improving robustness on hard samples with knots (such as T6_B3_N21_D), is a clear target.

Target Audience

• Computer vision researchers working on instance segmentation, boundary detection, or domain generalization who want a real-world annotated benchmark with 1930 ring annotations and cross-domain evaluation protocols.
• Forestry and wood-science researchers interested in quantifying the effect of silvicultural treatments on growth, and in moving from radial width measurements to 2D or volumetric ring metrics.
• Dendrochronologists and climate scientists who need fast, accurate ring delineation to correlate growth with temperature and precipitation records.
• Practitioners in forest management and the wood processing industry evaluating automated, image-based measurement pipelines; the dataset is released under CC BY 4.0 at the Zenodo repository (DOI 10.5281/zenodo.15653340).

Authors’ abstract

Tree-ring growth represents the annual wood increment for a tree, and quantifying it allows researchers to assess which silvicultural practices are best suited for each species. Manual measurement of this growth is time-consuming and often imprecise, as it is typically performed along 4 to 8 radial directions on a cross-sectional disc. In recent years, automated algorithms and datasets have emerged to enhance accuracy and automate the delineation of annual rings in cross-sectional images. To address the scarcity of wood cross-section data, we introduce the UruDendro4 dataset, a collection of 102 image samples of Pinus taeda L., each manually annotated with annual growth rings. Unlike existing public datasets, UruDendro4 includes samples extracted at multiple heights along the stem, allowing for the volumetric modeling of annual growth using manually delineated rings. This dataset (images and annotations) allows the development of volumetric models for annual wood estimation based on cross-sectional imagery. Additionally, we provide a performance baseline for automatic ring detection on this dataset using state-of-the-art methods. The highest performance was achieved by the DeepCS-TRD method, with a mean Average Precision of 0.838, a mean Average Recall of 0.782, and an Adapted Rand Error score of 0.084. A series of ablation experiments were conducted to empirically validate the final parameter configuration. Furthermore, we empirically demonstrate that training a learning model including this dataset improves the model's generalization in the tree-ring detection task.