A computer-vision study quantifying and mitigating the accuracy drop that occurs when a strawberry-ripeness classifier trained on daytime images is deployed under onboard LED lighting at night — a real failure mode for agricultural harvesting robots.
Autonomous strawberry-harvesting robots are typically trained on daytime field imagery but operate around the clock under onboard LED illumination. The LED spectral distribution shifts the input distribution away from the training data (domain shift), degrading ripeness-classification accuracy at night — exactly when a reliable model is most needed. This project measures that degradation rigorously and tests whether modern CNNs and domain-adaptation techniques close the gap.
Framed as patch classification (not object detection) on a public agricultural dataset (Pastell & Liakka, Luke Finland; 813 images, YOLO annotations, CC BY 4.0), with a fixed train/val/test split frozen up front and reused across every experiment for fair comparison. Five controlled experiments:
| Exp | What it tests |
|---|---|
| E1 | Deep-learning baseline (ConvNeXt) on daytime patches |
| E2 | Classical ML baseline (SVM on HSV colour features), day vs night |
| E3–E5 | Quantify the day→night ΔF1 and test domain-adaptation strategies |
Findings are interpreted with Grad-CAM to show where the models attend and how that attention breaks down under the lighting shift. Hypotheses (e.g. "a colour-sensitive HSV+SVM baseline degrades more than a deep model under domain shift") are stated in advance and tested empirically.
PyTorch · ConvNeXt (transfer learning) · scikit-learn (SVM) · Grad-CAM · HSV colour analysis · macro-F1 / per-class precision-recall · Google Colab (T4).
Reproducibility: fixed seeds across Python/NumPy/PyTorch(+CUDA),
cudnn.deterministic=True, and a persisted split CSV used as ground truth.
The headline finding: a classical colour-based model collapses under the night shift, while the deep model stays robust — confirming the central hypothesis.
| Experiment | Model | Test split | Macro-F1 | AUC-ROC |
|---|---|---|---|---|
| E1 | ConvNeXt-Tiny | day | 0.957 | 0.981 |
| E2 | SVM (RBF, HSV) | day | 0.905 | 0.967 |
| E2 | SVM (RBF, HSV) | night | 0.748 ⬇ | 0.891 |
| E3→E4 | ConvNeXt-Tiny | day → night | ΔF1 = 0.008 | — |
- The SVM baseline drops 0.905 → 0.748 (≈16 pts macro-F1) from day to night — a statistically significant degradation (95% CI), as predicted for a colour-sensitive model under LED-induced shift.
- ConvNeXt-Tiny degrades by only ΔF1 ≈ 0.008 day→night (E4) and retains 99.7% of its day-time performance — i.e. the deep model is largely shift-robust where the classical one is not.
Full per-class metrics, ROC/PR curves, confusion matrices, training curves, and
Grad-CAM visualisations are in results/ and the notebook
strawberry_domain_shift.ipynb.
# Open in Google Colab (GPU runtime) and run all cells:
# strawberry_domain_shift.ipynb
# The notebook mounts Google Drive for the dataset, checkpoints, and the
# frozen split CSV, then runs experiments E1–E5 in order.- Single public dataset; results would benefit from validation on additional field-captured night imagery.
- Patch classification simplifies the real detection-then-classify pipeline.
- Future work: explicit illumination-invariant augmentation / colour-constancy pre-processing, and a generative day→night style transfer for adaptation.
Individual MSc Applied AI coursework (University of Warwick, WMG — AI & Deep Learning).