Nwpu Resisc45 Eval

---
name: nwpu-resisc45-eval
description: Evaluates the ability of image classification models to accurately categorize remote sensing scenes into one of 45 predefined land-use/land-cover categories. It probes robustness to realistic variations in spatial resolution, viewpoint, illumination, occlusion, and object pose that are common in aerial imagery. Use when the user wants to benchmark on NWPU-RESISC45, or asks about evaluating this task. Reports overall accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 1703.00121
  bibtex_key: cheng2017remote
  confidence: high
---

# nwpu-resisc45-eval

> Remote Sensing Image Scene Classification: Benchmark and State of the Art — Cheng et al. (2017) (arXiv:1703.00121, 2017)

## What this evaluates

Evaluates the ability of image classification models to accurately categorize remote sensing scenes into one of 45 predefined land-use/land-cover categories. It probes robustness to realistic variations in spatial resolution, viewpoint, illumination, occlusion, and object pose that are common in aerial imagery.

## Datasets

- **NWPU-RESISC45** — total 31500; splits: train (-1), test (-1)

## Metrics

- `overall accuracy` **(primary)** — range: [0, 1]
  - The proportion of correctly classified test samples out of the total number of test samples. Calculated as (number of correct predictions) / (total test samples).

## Input / output format

**Input**: Remote sensing images of various geographic scenes.

**Output**: A single class label from the 45 predefined scene categories.

## Scoring recipe

```python
def compute_accuracy(predictions, labels):
    correct = sum(1 for p, l in zip(predictions, labels) if p == l)
    return correct / len(labels)

# Evaluation protocol:
# 1. Split dataset into train (10% or 20%) and test sets.
# 2. Train model on train split.
# 3. Predict labels for test set.
# 4. Compute overall accuracy using the function above.
# 5. Generate confusion matrix where entry (i,j) = rate of test samples from class i classified as class j.
```

## Common pitfalls

- The evaluation uses fixed random training ratios (10% and 20%) rather than a single deterministic train/test split, meaning results can vary across different random seeds.
- Confusion matrices are only reported for the highest-accuracy method within each feature category, not for all baselines, which hides per-class error patterns for weaker models.

## Evidence (verbatim from paper)

> Tables 3-6 show the overall accuracies of three handcrafted global features, three unsupervised feature learning methods, three deep CNN features, and three fine-tuned CNN features, respectively, under the training ratios of 10% and 20%.

## Citation

```bibtex
@misc{cheng2017remote,
  title={Remote Sensing Image Scene Classification: Benchmark and State of the Art},
  author={Cheng et al. (2017)},
  year={2017},
  note={arXiv:1703.00121}
}
```

- arXiv: 1703.00121