Beyond Candidate Coverage: 3D Point-Profile Re-scoring for Individual Tree Detection in Agroforestry

• Murilo Salem UFPel / SENAI-RS
• Luísa Böhm UFPel / SENAI-RS
• Henrique dos Reis UFPel
• Lúcio Vagner Carvalho UFPel
• Theo Ferrugem UFPel
• Diniz Arruda Carvalho SENAI-RS
• Anderson Ferrugem UFPel

DOI: https://doi.org/10.5753/eniac.2026.26892

Resumo

Individual tree detection in agroforestry systems is harder than in natural forests: dense, overlapping canopies generate spurious local maxima that are easily confused with real tree apices. We address this with a multimodal benchmark of 10 agroforestry plots (∼3,600 field-annotated trees) combining aerial RGB, 13-channel structural LiDAR, a canopy height model, and the raw 3D point cloud, evaluated under a leakage-free protocol (fixed validation/test plots, macro-averaged F1). We propose a two-stage decoupled detector: a convolutional network generates high-recall candidates, which an ExtraTrees classifier then re-scores using per-candidate engineered features, the vertical 3D profile of the raw point cloud, and a cross-detector agreement signal. The pipeline reaches F1 = 0.567 ± 0.004, above an end-to-end CNN ensemble (0.530; below the lower bound of an instance-level bootstrap confidence interval, though inside the wider between-plot band) and a classical watershed pipeline (0.352), raising recall from 0.50 to 0.60 while holding precision near 0.55. Critically, a controlled experiment shows that expanding candidate coverage to 99.8% of trees leaves F1 flat, and that none of the ∼25 re-scoring families we tested surpasses the same ceiling: randomized tree ensembles converge on it, while higher-capacity learners fall short of it. We conclude that the performance ceiling is largely a discriminability limit imposed by the available sensors and resolution—rather than by candidate coverage or model capacity—reframing future work toward richer spectral/full-waveform data and label-position review instead of larger models.

Referências

Balestra, M., Marselis, S., Sankey, T. T., Cabo, C., Liang, X., Mokroš, M., Peng, X., Singh, A., Stereńczak, K., Vega, C., Vincent, G., and Hollaus, M. (2024). LiDAR data fusion to improve forest attribute estimates: A review. Current Forestry Reports, 10:281–297.

Burges, C. J. C. (2010). From RankNet to LambdaRank to LambdaMART: An overview. Technical Report MSR-TR-2010-82, Microsoft Research.

Chen, Q., Baldocchi, D., Gong, P., and Kelly, M. (2006). Isolating individual trees in a savanna woodland using small footprint lidar data. Photogrammetric Engineering & Remote Sensing, 72(8):923–932.

Chen, T. and Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), pages 785–794.

Dalponte, M. and Coomes, D. A. (2016). Tree-centric mapping of forest carbon density from airborne laser scanning and hyperspectral data. Methods in Ecology and Evolution, 7(10):1236–1245.

Efron, B. and Tibshirani, R. J. (1993). An Introduction to the Bootstrap. Chapman & Hall/CRC, New York.

Geurts, P., Ernst, D., and Wehenkel, L. (2006). Extremely randomized trees. Machine Learning, 63(1):3–42.

Izmailov, P., Podoprikhin, D., Garipov, T., Vetrov, D., and Wilson, A. G. (2018). Averaging weights leads to wider optima and better generalization. In Proceedings of the Conference on Uncertainty in Artificial Intelligence (UAI).

Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., and Liu, T.-Y. (2017). LightGBM: A highly efficient gradient boosting decision tree. In Advances in Neural Information Processing Systems (NeurIPS), volume 30, pages 3146–3154.

Koch, B., Heyder, U., and Weinacker, H. (2006). Detection of individual tree crowns in airborne lidar data. Photogrammetric Engineering & Remote Sensing, 72(4):357–363.

Popescu, S. C. and Wynne, R. H. (2004). Seeing the trees in the forest: Using lidar and multispectral data fusion with local filtering and variable window size for estimating tree height. Photogrammetric Engineering & Remote Sensing, 70(5):589–604.

Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A. V., and Gulin, A. (2018). Cat-Boost: Unbiased boosting with categorical features. In Advances in Neural Information Processing Systems (NeurIPS), volume 31, pages 6638–6648.

Weinstein, B. G., Graves, S. J., Marconi, S., Singh, A., Zare, A., Stewart, D., Bohlman, S. A., and White, E. P. (2021). A benchmark dataset for canopy crown detection and delineation in co-registered airborne RGB, LiDAR and hyperspectral imagery from the national ecological observatory network. PLoS Computational Biology, 17(7):e1009180.

Weinstein, B. G., Marconi, S., Aubry-Kientz, M., Vincent, G., Senyondo, H., and White, E. P. (2020). DeepForest: A Python package for RGB deep learning tree crown delineation. Methods in Ecology and Evolution, 11(12):1743–1751.

Weinstein, B. G., Marconi, S., Bohlman, S., Zare, A., and White, E. (2019). Individual tree-crown detection in RGB imagery using semi-supervised deep learning neural networks. Remote Sensing, 11(11):1309.

Woo, S., Park, J., Lee, J.-Y., and Kweon, I. S. (2018). CBAM: Convolutional block attention module. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3–19.

Zhou, X., Wang, D., and Krähenbühl, P. (2019). Objects as points. arXiv preprint arXiv:1904.07850.