Reference-like Image-Level Dataset Split

Reference-like Image-Level Dataset Split This archive documents the procedure used to construct Dataset 2 (Reference-like) for the tree-species detection experiments. The purpose of this partition was to reproduce, as closely as possible, the relative training, validation, and test composition reported in the reference study while adapting those proportions to the number of images and annotations available in the present dataset. Partitioning strategy The split was generated at the image level rather than at the individual bounding-box level. All bounding boxes associated with the same source image were therefore kept in the same subset. This constraint prevents the same aerial image from contributing annotations to more than one of the training, validation, or test partitions. The class-wise and overall train/validation/test proportions reported in the reference study were first calculated and then scaled to the available dataset. These scaled values were used as the target image and class-annotation counts for the optimization procedure. The resulting target numbers of images were: Subset Images Train 3,707 Validation 412 Test 319 The corresponding target numbers of annotated trees were: Class Train Validation Test C1 — Norway spruce 2,714 302 433 C2 — Silver fir 2,007 224 120 C3 — Scots pine 511 57 25 C4 — European beech 3,229 359 147 The resulting partitions contained: Subset Images Annotated trees Train 3,707 8,461 Validation 412 942 Test 319 725 Optimization Image assignment was formulated as a mixed-integer linear programming (MILP) problem and solved using scipy.optimize.milp with the HiGHS optimization backend. The optimization was designed to: assign every image to exactly one subset; preserve all annotations belonging to the same image; satisfy the target image counts; reproduce the scaled class-specific targets as closely as possible; and prevent image overlap between training, validation, and test subsets. The recorded optimization completed successfully with an optimal HiGHS solution. A deterministic seed of 20260726 was used for the recorded partitioning procedure. Image-disjoint verification The final split was explicitly checked for image overlap. The recorded overlap counts were: Train vs. Validation: 0 Train vs. Test: 0 Validation vs. Test: 0 Thus, the final Dataset 2 configuration is image-disjoint. Files included The archive contains the following files: make_reference_like_split.pyRecords the Dataset 2 reference-like partition targets and the configuration used for reconstruction of the split. README_reference_like_split.txtProvides a concise human-readable description of the partitioning procedure and the final image- and annotation-level statistics. reference_like_split_metadata.jsonContains the detailed provenance of the split, including: counts reported in the reference study; scaled image targets; scaled class-specific targets; final partition statistics; image-overlap verification; optimization method; solver status; and deterministic seed. The original input annotation CSV files were treated as read-only during construction of the split and were not modified. Intended use These files are provided to document and support reproducibility of the Dataset 2 (Reference-like) configuration used in the accompanying tree-species object-detection experiments. Dataset 2 was used as a reference-like experimental partition for evaluating supervised Faster R-CNN and subsequent pseudo-label-based annotation expansion experiments. The archived split information should be used when reproducing the reported Dataset 2 experiments, since the exact membership of the training, validation, and test subsets is part of the experimental design. Reproducibility note The aggregate image and class targets describe the statistical structure of the Dataset 2 partition. For exact reproduction of the experiments, the final image-level split files or manifests distributed with the dataset should be regarded as the authoritative record of subset membership. The optimization metadata is included to document how the reference-like partition was derived and to allow the same partitioning strategy to be applied to equivalent annotation datasets.