Data and code for risk ranking and threshold transfer in delayed image plane crop targeting
Description
Research code and numerical data for the manuscript Risk ranking and threshold transfer for delayed image plane crop targeting. Real static PhenoBench sugar-beet images supply crop detections; motion, perception delay and execution errors are explicitly simulated in the image plane. This is an offline evaluation, not a hardware or field-control trial. The fixed protocol uses three YOLOv8n models (seeds 17, 29 and 43; 40 training epochs each), 219 evaluation images with 1190 eligible crop observations, five simulated conditions and 20 motion replicates. The archive includes the fixed split and protocol, converted annotations, saved predictions, training settings and logs, simulated candidate events, calibrated thresholds, exported scoring models, complete result tables, plotting code and figures. The primary Kaggle CPU run and the separate macOS environment sensitivity are retained separately. Logistic-regression and full-candidate-pool comparisons are explicitly post hoc. README.md gives reproduction commands; DATA_DICTIONARY.md explains units and denominators; SHA256SUMS.json and verify_deposit.py support integrity checking. Raw PhenoBench images/masks, detector weights, third-party article PDFs and internal reviewer reports are not redistributed. Obtain source images and software from their providers. The source paper is Weyler et al., IEEE TPAMI 46, 9583-9594 (2024), doi:10.1109/TPAMI.2024.3419548. The mirror used for the study identifies CC BY-NC-SA 4.0, while the current official dataset pages identify CC BY-SA 4.0. Original notices and this discrepancy are retained; the deposit follows the mirror's noncommercial share-alike terms. See LICENSE_SCOPE.md. OpenAI Codex assisted with code, numerical checks, explanatory diagrams and manuscript preparation. No generated crop photographs are represented as observations. Ranking improvements are conditional: comparisons reverse under nominal delay and at 32 pixels, and improved ranking does not guarantee the reliability of a frozen threshold. Motion replicates are conditional simulations, not independent field experiments. Author affiliations: School of Blockchain, Jiangxi University of Software Professional Technology, Nanchang 330041, China; School of Engineering and Technology, Shinawatra University, Pathum Thani 12160, Thailand.
Files
Steps to reproduce
1. Download and extract AUTOAI002_data_code_v1.zip. Work from the extracted AUTOAI002_data_code_v1 directory and keep newly generated outputs outside it. Run python3 -B verify_deposit.py . to check exact file names, sizes and SHA256 hashes. 2. Create a Python 3.12 environment and install 06_Reproducibility/AUTOAI002/requirements-simulation.txt. The primary simulation is 04_Results/AUTOAI002/kaggle_simulation_v001/autoai002_simulation; the separate local_simulation_v001 folder is the macOS sensitivity run. 3. Replay saved events with 03_Experiments/AUTOAI002/audit_results.py RESULT_DIRECTORY --report OUTPUT.json. Use audit_saved_models.py with the same arguments to evaluate exported models and calibration choices. Run audit_supplementary.py --report OUTPUT.json for the post hoc checks. Full commands and their validation limits are in README.md. 4. To refit the scoring models and rerun the full simulation from archived detections, run simulation_run.py --vision-root 04_Results/AUTOAI002/vision_v002/autoai002_vision --out NEW_OUTPUT_DIRECTORY. Detector weights are not required for this path. Platform and library differences can affect fits and thresholds; the saved Kaggle run remains the source of primary manuscript values. 5. Rebuild tables with analyse_results.py, specifying --raw, --local-results and --out as documented. Install requirements-figures.txt and run plot_results.py with --data and --out. The graphical-abstract builder uses ReportLab and explicit font-file arguments; see README.md. 6. Repeating detector training requires separately obtaining PhenoBench 1.1.0 and compatible GPU resources. The original kaggle_vision/experiment.ipynb embeds the training script, protocol and split. Source image and model hashes and the recorded environments are retained. Original Kaggle run identifiers refer to private runs and are not public access links. The root checksum list describes this exact deposit. The original visual source manifest, including hashes of intentionally omitted images and weights, is preserved separately for provenance. A reduced visual-folder checksum list validates only retained inputs. All primary and macOS simulation result files are unchanged from the frozen research package.