Add per-timestamp NDVI columns (``t{t}_ndvi``) to *df*. NDVI = (NIR - Red) / (NIR + Red). A small epsilon avoids division by zero.

Return ``(X, y)`` where X has shape ``(n, n_timestamps * features_per_step)``. Feature order within each timestep: [b0, b1, …, b{N_BANDS-1},

1-D CNN over the temporal axis with global average pooling.

Return a DataFrame with shape (n_classes * n_samples, n_timestamps*n_bands + 1). Columns are named ``t{t}_b{b}`` for timestamp t, band b, pl

Bidirectional LSTM — the recommended architecture for temporal spectral data. Processing the sequence in both directions captures phenologic

Fully-connected baseline operating on the flat feature vector.

Print and save: - Per-class precision / recall / F1 report - Confusion matrix PNG - Training curve PNG (if *history* passed separatel

Save loss and accuracy curves from a Keras History object.

Train a crop-classification model and return evaluation artefacts. Parameters ---------- df : pd.DataFrame Output of ``prepr

Dataset must have N_TIMESTAMPS * N_BANDS band columns + 1 class column.

Flat array must reshape cleanly to (n, T, F) without reordering data.

Spot-check NDVI = (NIR - Red) / (NIR + Red) for a known pixel.

Each class block should have unique pixel locations (no replacement).

Call build_dataset with mocked GDAL and assert that _load_image is called with the *same* row/col index arrays for every timestamp.

When n_samples > available pixels, all available pixels are used.