\

FOCR1 longitudinal hyperspectral analysis

Reproducible Python implementation of the analysis underlying:

From Spectra to Signal: Temporal Detection of Fusarium Wilt in Banana with Block-Aware Machine Learning

Scope

This repository reproduces the paper analysis from the preserved curated dataset forward. It covers final QC, the E5 Block 7 sensitivity decision, block-aware temporal inference, leaf/position analysis, exploratory spectral-region identification, leave-one-block-out machine learning, and QC sensitivity.

The original one-off analysis scripts were not preserved in the exported analysis packages. This repository is therefore a reconstructed reproducible implementation, validated against the frozen numerical results. It should not be described as a byte-for-byte copy of the original transient code.

Input

Copy the file below from 00. FOCR1_curated_data_package:

banana_spectral_master_clean.csv

to:

data/banana_spectral_master_clean.csv

The curated input contains 2,880 expected measurement slots, including 2,873 observed spectra and seven missing measurements. It also preserves traceable metadata/ID corrections, three reconstructed split spectra, and the low-signal QC flag.

Install

python -m venv .venv
source .venv/bin/activate      # Windows: .venv\Scripts\activate
pip install -r requirements.txt

Reproduce all analyses

python run_pipeline.py

or run stages separately:

python src/00_validate_curated.py
python src/01_qc_final.py
python src/02_exploration.py
python src/03_e5_sensitivity.py
python src/04_temporal_separability.py
python src/05_leaf_position.py
python src/06_spectral_regions.py
python src/07_machine_learning.py
python src/08_qc_sensitivity.py
python src/09_validate_paper_results.py

Fixed machine-learning configuration

No hyperparameter tuning or nested cross-validation was used. Hyperparameters are fixed before leave-one-block-out evaluation; no held-out block is used for tuning.

Model	Fixed configuration
Logistic regression	L2, C=1.0, solver="liblinear", class_weight="balanced"
RBF-SVM	C=1.0, gamma="scale", class_weight="balanced"
PLS-DA	5 latent components; class threshold 0.5

All standardization parameters are estimated only on the training blocks in each fold.

These settings were recovered by reconstructing the preserved LOBO pipeline and matching the archived out-of-fold prediction scores. In particular, the archived logistic scores match liblinear, C=1, the SVM scores match C=1, gamma="scale", and the PLS-DA scores match five latent components.

Frozen checks

src/09_validate_paper_results.py verifies the principal paper claims, including:

• 2,761 spectra after final QC;
• 2,703 spectra after excluding E5 Block 7;
• Foc R1-only E4 VIS mean difference ≈ +0.00441;
• exact block sign-flip p = 0.0078125;
• Holm-adjusted p = 0.0390625;
• E4 whole-plant 520–749 nm RBF-SVM balanced accuracy ≈ 0.703;
• E5 whole-plant full-spectrum logistic balanced accuracy ≈ 0.799.

Run unit checks with:

pytest -q

Statistical design

The independently randomized unit is the plant, not an individual spectrum. Repeated leaf/position spectra are therefore never treated as independent plants.

Temporal confirmatory inference:

• plant-level averaging within evaluation;
• treatment-control contrasts computed within randomized blocks;
• exact two-sided sign-flip tests;
• Holm correction across five predefined regions;
• exact maxT family-wise correction for bandwise inference.

Predictive evaluation:

• leave one complete experimental block out at each fold;
• preprocessing/scaling fitted on training blocks only;
• predictions pooled across held-out blocks.

Important terminology

The labels represent experimental inoculation status, not independently confirmed infection status. Code and outputs use FOC_R1_only, Radopholus_only, Coinoculation, and Control as treatment-group labels.

Data availability

The repository does not make a claim about public availability of the biological data. Update this section only after the authors/data owners decide whether the dataset can be public, available on request, or restricted.