Fujiview Svf Eval

---
name: fujiview-svf-eval
description: Evaluates multimodal late-fusion models for predicting scenic visibility (clear, cloudy, perfect, obscured) across short- to medium-term forecasting horizons (+0d to +3d). It probes the model's ability to integrate visual webcam features with meteorological forecasts to handle class imbalance and temporal dynamics in environmental perception. Use when the user wants to benchmark on FujiView, or asks about evaluating this task. Reports accuracy (ACC).
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2603.00157
  bibtex_key: bible2026fujiview
  confidence: high
---

# fujiview-svf-eval

> FujiView: Multimodal Late-Fusion for Predicting Scenic Visibility — Bible et al. (2026) (arXiv:2603.00157, 2026)

## What this evaluates

Evaluates multimodal late-fusion models for predicting scenic visibility (clear, cloudy, perfect, obscured) across short- to medium-term forecasting horizons (+0d to +3d). It probes the model's ability to integrate visual webcam features with meteorological forecasts to handle class imbalance and temporal dynamics in environmental perception.

## Datasets

- **FujiView** — total 100000; splits: train (-1), val (-1), test (-1)

## Metrics

- `accuracy (ACC)` **(primary)** — range: percent
  - Percentage of correct predictions out of total instances. Computed as the proportion of samples where the predicted visibility class matches the ground truth label.
- `AUC` — range: [0, 1]
  - Area under the Receiver Operating Characteristic curve, measuring the model's ability to rank positive instances higher than negative ones across all classification thresholds.

## Input / output format

**Input**: Concatenated feature vector containing calibrated softmax probabilities from a YOLOv8n-cls vision model (trained on webcam images) and meteorological features (current conditions and forecasts up to 3 days ahead).

**Output**: Predicted visibility class label (OBSCURED, PERFECT, CLEAR, or CLOUDY).

## Scoring recipe

```python
def compute_metrics(predictions, labels, probabilities):
    acc = sum(1 for p, l in zip(predictions, labels) if p == l) / len(labels)
    auc = roc_auc_score(labels, probabilities, multi_class='ovr')
    return acc, auc
```

## Common pitfalls

- Class imbalance heavily favors the OBSCURED class, which can bias models toward predicting non-visibility if not handled.
- Horizon difficulty is non-monotonic; +1d forecasting can be easier than +0d due to synoptic weather stability vs. transient cloud dynamics.
- Temporal windowing (averaging frames over 3 hours) improves vision-only AUC but may oversmooth features and reduce accuracy for fusion models at longer horizons.

## Evidence (verbatim from paper)

> Table 1 summarizes mean accuracy (ACC)—the percentage of correct predictions—and area under the ROC curve (AUC)—a threshold-independent measure of ranking performance—computed from 5-fold grouped cross-validation.

## Citation

```bibtex
@misc{bible2026fujiview,
  title={FujiView: Multimodal Late-Fusion for Predicting Scenic Visibility},
  author={Bible et al. (2026)},
  year={2026},
  note={arXiv:2603.00157}
}
```

- arXiv: 2603.00157