SelvaBox: A high‑resolution dataset for tropical tree crown detection

ICLR 2026 Conference SubmissionAnonymous Authors
Remote sensingForest monitoringTree crown detectionTropical forest dataset
Abstract:

Detecting individual tree crowns in tropical forests is essential to study these complex and crucial ecosystems impacted by human interventions and climate change. However, tropical crowns vary widely in size, structure, and pattern and are largely overlapping and intertwined, requiring advanced remote sensing methods applied to high-resolution imagery. Despite growing interest in tropical tree crown detection, annotated datasets remain scarce, hindering robust model development. We introduce SelvaBox, the largest open‑access dataset for tropical tree crown detection in high-resolution drone imagery. It spans three countries and contains more than 8300083\,000 manually labeled crowns -- an order of magnitude larger than all previous tropical forest datasets combined. Extensive benchmarks on SelvaBox reveal two key findings: 1) higher-resolution inputs consistently boost detection accuracy; and 2) models trained exclusively on SelvaBox achieve competitive zero-shot detection performance on unseen tropical tree crown datasets, matching or exceeding competing methods. Furthermore, jointly training on SelvaBox and three other datasets at resolutions from 3 to 10 cm per pixel within a unified multi-resolution pipeline yields a detector ranking first or second across all evaluated datasets. Our dataset, code, and pre-trained weights are made public.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces SelvaBox, a large-scale annotated dataset containing over 83,000 manually labeled tropical tree crowns across three countries, positioned within the Training Data and Benchmark Datasets leaf of the taxonomy. This leaf contains only two papers, indicating a relatively sparse research direction focused on curating standardized resources for model development. The dataset contribution addresses a recognized bottleneck in tropical forest remote sensing, where annotated imagery remains scarce despite growing interest in automated crown detection methods.

The taxonomy reveals that SelvaBox sits within the broader Deep Learning-Based Detection Methods branch, which encompasses specialized architectures (one-stage detectors, instance segmentation networks), foundation model adaptations (SAM-based approaches, parameter-efficient fine-tuning), and active learning strategies. Neighboring branches include LiDAR-Based Detection and Photogrammetric Structure-from-Motion Approaches, which leverage three-dimensional data rather than RGB imagery alone. The dataset's multi-country scope and high-resolution focus distinguish it from species-specific detection efforts (e.g., palm identification) and ecological monitoring applications that track temporal dynamics rather than static crown delineation.

Among the 30 candidates examined, none clearly refute the three core contributions: the SelvaBox dataset itself (10 candidates examined, 0 refutable), the benchmark evaluation framework (10 examined, 0 refutable), and the multi-resolution training pipeline (10 examined, 0 refutable). The sibling paper ReforesTree addresses restoration contexts with potentially different annotation strategies, but the limited search scope prevents definitive claims about overlap. The statistics suggest that within the examined literature, no prior work provides a tropical crown dataset of comparable scale or geographic diversity, though the analysis does not cover exhaustive domain-specific repositories or unpublished institutional datasets.

Based on the top-30 semantic matches and taxonomy structure, the work appears to occupy a relatively underserved niche in tropical forest benchmarking. The sparse population of the Training Data leaf and absence of refuting candidates among examined papers suggest novelty in dataset scale and multi-country coverage. However, the limited search scope means the analysis cannot rule out overlapping efforts in specialized forestry journals, regional datasets, or concurrent preprints not captured by the semantic search strategy.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
30
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: tropical tree crown detection in high-resolution drone imagery. The field has evolved into a multi-faceted landscape organized around several complementary branches. Deep Learning-Based Detection Methods form a central pillar, encompassing works that develop neural architectures and training strategies for automated crown delineation, such as MTCDNet[4] and TCDNet[10]. LiDAR-Based Detection and Segmentation leverages three-dimensional point clouds to capture canopy structure, while Photogrammetric Structure-from-Motion Approaches reconstruct forest geometry from overlapping RGB imagery. Species-Specific Detection and Classification targets individual taxa or functional groups—illustrated by studies on palms like PalmProbNet[15] and Amazonian Palms Detection[8]—whereas Ecological and Phenological Monitoring Applications track temporal dynamics such as flowering events (Flowering Monitoring Drone[5]) and leaf phenology (Leaf Phenology UAV[1]). Biophysical Parameter Estimation and Scaling translates crown-level detections into forest-wide metrics like biomass, Forest Composition and Diversity Analysis quantifies community structure, and Operational Workflows and Field Integration addresses practical deployment challenges including flight planning and data processing pipelines. Within the deep learning branch, a particularly active line of work focuses on Training Data and Benchmark Datasets, recognizing that model performance hinges on the availability of high-quality annotated imagery. SelvaBox[0] contributes directly to this effort by providing a curated benchmark dataset designed to support reproducible evaluation and method comparison. This emphasis on standardized training resources contrasts with studies that prioritize architectural innovation (Deep Learning Ensemble[3]) or parameter-efficient adaptations (Parameter Efficient Tree Detection[2]). Meanwhile, ReforesTree[9]—a close neighbor in the taxonomy—also addresses data challenges but targets restoration contexts where labeled examples may be scarce. The interplay between dataset curation, model design, and domain-specific constraints remains a central open question: as tropical forests exhibit extreme structural and spectral heterogeneity, the community continues to explore whether general-purpose benchmarks or task-specific datasets better advance detection accuracy and ecological relevance.

Claimed Contributions

SelvaBox dataset for tropical tree crown detection

The authors present SelvaBox, a large-scale dataset comprising over 83,000 manually annotated tree crown bounding boxes from high-resolution drone imagery across Brazil, Ecuador, and Panama. This dataset is an order of magnitude larger than all previous tropical forest datasets combined and addresses the critical scarcity of annotated tropical forest data.

10 retrieved papers
Comprehensive benchmark and evaluation framework

The authors provide extensive benchmarks comparing CNN-based and transformer-based detection methods across multiple resolutions and input sizes. They introduce a raster-level evaluation metric (RF175) that addresses limitations of tile-level metrics and enables proper assessment of model performance on entire rasters rather than individual tiles.

10 retrieved papers
Multi-resolution training pipeline and state-of-the-art models

The authors develop a multi-resolution training approach that enables models to generalize across different spatial resolutions and ground sampling distances. Their trained models achieve state-of-the-art performance on both tropical and non-tropical datasets, with competitive zero-shot detection on unseen datasets, and they release pre-trained weights and two open-source Python libraries for preprocessing and benchmarking.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

SelvaBox dataset for tropical tree crown detection

The authors present SelvaBox, a large-scale dataset comprising over 83,000 manually annotated tree crown bounding boxes from high-resolution drone imagery across Brazil, Ecuador, and Panama. This dataset is an order of magnitude larger than all previous tropical forest datasets combined and addresses the critical scarcity of annotated tropical forest data.

Contribution

Comprehensive benchmark and evaluation framework

The authors provide extensive benchmarks comparing CNN-based and transformer-based detection methods across multiple resolutions and input sizes. They introduce a raster-level evaluation metric (RF175) that addresses limitations of tile-level metrics and enables proper assessment of model performance on entire rasters rather than individual tiles.

Contribution

Multi-resolution training pipeline and state-of-the-art models

The authors develop a multi-resolution training approach that enables models to generalize across different spatial resolutions and ground sampling distances. Their trained models achieve state-of-the-art performance on both tropical and non-tropical datasets, with competitive zero-shot detection on unseen datasets, and they release pre-trained weights and two open-source Python libraries for preprocessing and benchmarking.

SelvaBox: A high‑resolution dataset for tropical tree crown detection | Novelty Validation