# Tondevrel Scientific Agent Skills Rasterio

> Rasterio — Raster Geospatial Processing

- Skill: `tomevault-io/tondevrel-scientific-agent-skills-rasterio` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add tomevault-io/tondevrel-scientific-agent-skills-rasterio`
- Raw SKILL.md: https://api.skillmd.com/api/skills/tomevault-io/tondevrel-scientific-agent-skills-rasterio/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: tomevault-io (https://skillmd.com/u/tomevault-io)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/tomevault-io/tondevrel-scientific-agent-skills-rasterio

---


# Rasterio — Raster Geospatial Processing

Rasterio is the standard Python library for reading and writing georeferenced raster data. It wraps GDAL but exposes a clean Pythonic API. Every raster has two components: **pixel values** (a numpy array) and **geospatial metadata** (CRS, transform, bounds) that maps pixels to real-world coordinates.

## Raster vs Vector — When to Use What

```
RASTER (Rasterio)                      VECTOR (GeoPandas / Shapely)
────────────────────────────────       ────────────────────────────────
Grid of pixels                         Points, lines, polygons
Satellite imagery                      Administrative boundaries
Elevation models (DEM)                 Road networks
Land cover maps                        Locations / GPS tracks
Temperature grids                      Census regions
Spectral data                          Feature geometries

KEY RULE: When your data IS a grid → Rasterio.
          When your data IS shapes/points → GeoPandas.
          When you need BOTH → Rasterio clips rasters TO vector shapes.
```

## Reference Documentation

**Rasterio docs**: https://rasterio.readthedocs.io/en/latest/  
**GDAL formats**: https://gdal.org/drivers/raster/index.html  
**GitHub**: https://github.com/rasterio/rasterio  
**Search patterns**: `rasterio.open`, `dataset.read`, `dataset.transform`, `rasterio.mask`, `show`

## Core Principles

### The Dataset Object
`rasterio.open()` returns a dataset. A dataset has **bands** (layers of pixel data), a **transform** (maps pixel coordinates to geographic coordinates), a **CRS** (coordinate reference system), and **bounds** (geographic extent). Always use context manager (`with` statement) — it handles file handles correctly.

### The Transform
An affine transform maps pixel (col, row) to geographic (x, y). `dataset.transform` gives this mapping. `dataset.index(x, y)` gives the reverse: geographic → pixel. This is how you go from "latitude/longitude" to "which pixel?"

### Bands
Most satellite imagery is multi-band: Band 1 = Red, Band 2 = Green, Band 3 = Blue, Band 4 = NIR, etc. Band numbering starts at **1** (not 0). `dataset.read(1)` reads band 1 as a 2D numpy array.

### CRS — Coordinate Reference Systems
Every raster is projected into some CRS. `EPSG:4326` = WGS84 lat/lon (GPS coordinates). `EPSG:32633` = UTM zone 33N (meters, good for Europe). Operations between rasters in different CRS require reprojection first.

### NoData
Pixels outside valid coverage are marked with a `nodata` value (e.g., -9999, 0, or NaN). Always mask these before computation — including them corrupts statistics.

## Quick Reference

### Installation

```bash
pip install rasterio numpy matplotlib
# For vector integration:
pip install geopandas shapely fiona
```

### Standard Imports

```python
import rasterio
from rasterio.transform import from_bounds, Affine
from rasterio.crs import CRS
import numpy as np
import matplotlib.pyplot as plt
```

### Basic Pattern — Read, Inspect, Visualize

```python
import rasterio
import numpy as np
import matplotlib.pyplot as plt

with rasterio.open('image.tif') as src:
    # Metadata
    print(f"Bands:      {src.count}")
    print(f"Shape:      {src.height} x {src.width}")
    print(f"CRS:        {src.crs}")
    print(f"Bounds:     {src.bounds}")           # (left, bottom, right, top)
    print(f"Resolution: {src.res}")              # (x_res, y_res) in CRS units
    print(f"NoData:     {src.nodata}")
    print(f"Transform:  {src.transform}")

    # Read all bands → shape: (bands, height, width)
    data = src.read()

    # Read single band → shape: (height, width)
    band1 = src.read(1)

    # Read with masked nodata → numpy masked array
    band1_masked = src.read(1, masked=True)

# Visualize single band
plt.imshow(band1, cmap='viridis')
plt.colorbar(label='Value')
plt.title('Band 1')
plt.show()
```

### Basic Pattern — Write a Raster

```python
import rasterio
from rasterio.transform import from_bounds
import numpy as np

# Create a 100x100 single-band raster covering a geographic extent
height, width = 100, 100
data = np.random.rand(height, width).astype(np.float32)

transform = from_bounds(
    west=10.0, south=50.0, east=11.0, north=51.0,
    width=width, height=height
)

with rasterio.open(
    'output.tif',
    'w',
    driver='GTiff',
    height=height,
    width=width,
    count=1,                          # Number of bands
    dtype=data.dtype,
    crs='EPSG:4326',                  # WGS84
    transform=transform,
    nodata=-9999
) as dst:
    dst.write(data, 1)                # Write to band 1
```

## Critical Rules

### ✅ DO

- **Always use `with rasterio.open(...)` context manager** — Ensures file handles are closed properly. Never do `src = rasterio.open(...)` without `with`.
- **Use `masked=True` when reading** — Returns a numpy masked array that automatically excludes nodata pixels. Prevents nodata from corrupting calculations.
- **Match CRS before any spatial operation** — If two rasters have different CRS, reproject one before combining. Use `rasterio.warp.reproject()`.
- **Use windowed reading for large files** — Files > 1GB will OOM if read entirely. Use `src.read(window=...)` to read chunks.
- **Preserve geospatial metadata when writing** — Copy `transform`, `crs`, `nodata` from the source when creating derived rasters.
- **Use `float32` for computed indices** — NDVI, NDWI produce float values [-1, 1]. Input bands are often `uint8` or `uint16` — cast before division to avoid integer truncation.
- **Check `src.nodata` before computing** — If nodata is None, the file has no declared nodata value. Handle accordingly.
- **Use `rasterio.features` for vector-raster conversion** — Don't roll your own rasterization.

### ❌ DON'T

- **Don't mix up band indexing** — Rasterio bands start at **1**. `src.read(1)` = first band. numpy arrays are 0-indexed, so `data[0]` = first band after `src.read()`.
- **Don't forget to cast dtypes before band math** — `uint8 / uint8` = integer division = truncated to 0. Always cast: `band.astype(np.float32)`.
- **Don't assume all rasters share the same CRS** — Even "standard" datasets may differ. Always check and align.
- **Don't ignore resolution differences** — Two rasters covering the same area may have different pixel sizes. Resample to match before pixel-wise operations.
- **Don't hardcode nodata as 0** — Nodata values vary by file. Always read `src.nodata`.
- **Don't read entire multi-GB files into memory** — Use windowed I/O or overview levels.

## Anti-Patterns (NEVER)

```python
import rasterio
import numpy as np

# ❌ BAD: Integer division in band math — truncates to 0
with rasterio.open('sentinel.tif') as src:
    nir  = src.read(4)    # uint16
    red  = src.read(3)    # uint16
    ndvi = (nir - red) / (nir + red)   # INTEGER DIVISION → all zeros or ±1!

# ✅ GOOD: Cast to float BEFORE any arithmetic
with rasterio.open('sentinel.tif') as src:
    nir  = src.read(4).astype(np.float32)
    red  = src.read(3).astype(np.float32)
    ndvi = (nir - red) / (nir + red + 1e-10)   # +1e-10 avoids div-by-zero

# ─────────────────────────────────────────────────────────────

# ❌ BAD: No context manager — file handle leak
src  = rasterio.open('image.tif')
data = src.read(1)
# src never closed → resource leak, potential corruption on write

# ✅ GOOD: Context manager
with rasterio.open('image.tif') as src:
    data = src.read(1)
# File closed automatically

# ─────────────────────────────────────────────────────────────

# ❌ BAD: Ignoring nodata in statistics
with rasterio.open('dem.tif') as src:
    elev = src.read(1)
    mean_elevation = elev.mean()    # Includes -9999 nodata → wrong answer!

# ✅ GOOD: Mask nodata before computing
with rasterio.open('dem.tif') as src:
    elev   = src.read(1, masked=True)           # Masked array
    mean_elevation = float(elev.mean())          # Ignores masked (nodata) pixels
    # OR explicitly:
    valid = elev[elev != src.nodata]
    mean_elevation = valid.mean()

# ─────────────────────────────────────────────────────────────

# ❌ BAD: Operating on rasters with different CRS without reprojection
with rasterio.open('landcover_utm.tif') as src1:
    with rasterio.open('population_wgs84.tif') as src2:
        # src1.crs = EPSG:32633 (UTM), src2.crs = EPSG:4326 (WGS84)
        pop  = src2.read(1)
        land = src1.read(1)
        result = pop * land   # WRONG — pixels don't align geographically!

# ✅ GOOD: Reproject to common CRS first (see Reprojection section)
```

## Reading Rasters

### Full Read vs Selective Read

```python
import rasterio
import numpy as np

with rasterio.open('multispectral.tif') as src:
    # All bands at once → shape: (n_bands, height, width)
    all_bands = src.read()

    # Single band → shape: (height, width)
    band1 = src.read(1)

    # Multiple specific bands → shape: (n_selected, height, width)
    rgb = src.read([1, 2, 3])

    # With nodata masking
    band1_masked = src.read(1, masked=True)   # numpy.ma.MaskedArray

    # Specific spatial subset (window)
    from rasterio.windows import Window
    window = Window(col_off=100, row_off=200, width=50, height=50)
    patch  = src.read(1, window=window)       # shape: (50, 50)

    # Read at lower resolution (overview)
    # out_shape forces resampling to smaller array
    small = src.read(1, out_shape=(src.height // 4, src.width // 4))
```

### Coordinate ↔ Pixel Mapping

```python
import rasterio

with rasterio.open('image.tif') as src:
    # Geographic coordinates → pixel (row, col)
    row, col = src.index(lon, lat)       # e.g., src.index(10.5, 50.3)

    # Pixel (row, col) → geographic coordinates (x, y)
    x, y = src.xy(row, col)             # Center of that pixel

    # Geographic extent of a specific pixel
    # Top-left corner of pixel (row, col):
    x_tl = src.transform.c + col * src.transform.a
    y_tl = src.transform.f + row * src.transform.e

    # Bounds of the entire raster
    print(f"Left={src.bounds.left}, Bottom={src.bounds.bottom}, "
          f"Right={src.bounds.right}, Top={src.bounds.top}")
```

## Writing Rasters

### Write with Full Metadata

```python
import rasterio
from rasterio.transform import from_bounds, Affine
import numpy as np

def write_raster(data: np.ndarray,
                 output_path: str,
                 transform: Affine,
                 crs: str = 'EPSG:4326',
                 nodata: float = -9999,
                 band_descriptions: list[str] = None):
    """
    Write a numpy array as a georeferenced GeoTIFF.
    data shape: (bands, height, width) or (height, width) for single band.
    """
    if data.ndim == 2:
        data = data[np.newaxis, :]       # Add band dimension → (1, H, W)

    n_bands, height, width = data.shape

    with rasterio.open(
        output_path,
        'w',
        driver='GTiff',
        height=height,
        width=width,
        count=n_bands,
        dtype=data.dtype,
        crs=crs,
        transform=transform,
        nodata=nodata,
        compress='lzw',                  # Compression — reduces file size significantly
        tiled=True,                      # Tiled layout — better for large file random access
        blockxsize=256,
        blockysize=256
    ) as dst:
        dst.write(data)
        if band_descriptions:
            for i, desc in enumerate(band_descriptions, 1):
                dst.set_band_description(i, desc)

# ─── Copy metadata pattern (most common: derive new raster from existing) ───
with rasterio.open('input.tif') as src:
    computed = src.read(1).astype(np.float32) * 2.0   # Some computation

    # Copy ALL metadata from source, override only what changed
    profile = src.profile.copy()
    profile.update(dtype=computed.dtype, count=1, nodata=-9999)

    with rasterio.open('output.tif', 'w', **profile) as dst:
        dst.write(computed, 1)
```

## CRS and Reprojection

### Check and Compare CRS

```python
import rasterio
from rasterio.crs import CRS

with rasterio.open('image.tif') as src:
    print(f"CRS: {src.crs}")
    print(f"EPSG: {src.crs.to_epsg()}")
    print(f"WKT:  {src.crs.to_wkt()}")

    # Check if two CRS are the same
    target_crs = CRS.from_epsg(32633)   # UTM zone 33N
    print(f"Same as UTM 33N? {src.crs == target_crs}")

# Common EPSG codes:
# EPSG:4326  — WGS84 lat/lon (GPS, most web maps)
# EPSG:3857  — Web Mercator (Google Maps, OpenStreetMap tiles)
# EPSG:32601–32660 — UTM zones 1–60 North (meters, good for local analysis)
# EPSG:32701–32760 — UTM zones 1–60 South
```

### Reproject a Raster

```python
import rasterio
from rasterio.warp import calculate_default_transform, reproject, Resampling
from rasterio.crs import CRS
import numpy as np

def reproject_raster(input_path: str, output_path: str, target_crs: str = 'EPSG:4326'):
    """Reproject an entire raster file to a new CRS."""
    target_crs = CRS.from_user_input(target_crs)

    with rasterio.open(input_path) as src:
        # Calculate the optimal transform and dimensions for the target CRS
        transform, width, height = calculate_default_transform(
            src.crs, target_crs,
            src.width, src.height,
            *src.bounds
        )

        # Prepare output profile
        profile = src.profile.copy()
        profile.update(crs=target_crs, transform=transform, width=width, height=height)

        with rasterio.open(output_path, 'w', **profile) as dst:
            for band_idx in range(1, src.count + 1):
                reproject(
                    source=rasterio.band(src, band_idx),
                    destination=rasterio.band(dst, band_idx),
                    src_transform=src.transform,
                    src_crs=src.crs,
                    dst_transform=transform,
                    dst_crs=target_crs,
                    resampling=Resampling.bilinear   # bilinear for continuous data
                    # Use Resampling.nearest for categorical (land cover class IDs)
                )

# reproject_raster('utm_image.tif', 'wgs84_image.tif', 'EPSG:4326')
```

## Band Math — Spectral Indices

```python
import rasterio
import numpy as np

def compute_indices(input_path: str, output_path: str,
                    band_map: dict = None) -> dict:
    """
    Compute standard spectral indices from multispectral imagery.

    band_map: maps index name to band number.
    Sentinel-2 default: {'blue': 2, 'green': 3, 'red': 4, 'nir': 8, 'swir1': 11, 'swir2': 12}
    Landsat 8 default:  {'blue': 2, 'green': 3, 'red': 4, 'nir': 5, 'swir1': 6, 'swir2': 7}
    """
    if band_map is None:
        band_map = {'blue': 2, 'green': 3, 'red': 4, 'nir': 8, 'swir1': 11, 'swir2': 12}

    with rasterio.open(input_path) as src:
        # Read required bands as float32
        bands = {}
        for name, idx in band_map.items():
            bands[name] = src.read(idx, masked=True).astype(np.float32)

        eps = 1e-10   # Avoid division by zero

        # ─── NDVI: Normalized Difference Vegetation Index ───
        # Range: [-1, 1]. Vegetation > 0.3. Bare soil ≈ 0. Water < 0.
        ndvi = (bands['nir'] - bands['red']) / (bands['nir'] + bands['red'] + eps)

        # ─── NDWI: Normalized Difference Water Index ───
        # Range: [-1, 1]. Water > 0. Vegetation < 0.
        ndwi = (bands['green'] - bands['nir']) / (bands['green'] + bands['nir'] + eps)

        # ─── NDSI: Normalized Difference Snow Index ───
        ndsi = (bands['green'] - bands['swir1']) / (bands['green'] + bands['swir1'] + eps)

        # ─── EVI: Enhanced Vegetation Index ───
        # More robust than NDVI in high-biomass areas
        evi = 2.5 * (bands['nir'] - bands['red']) / (
              bands['nir'] + 6 * bands['red'] - 7.5 * bands['blue'] + 1 + eps)

        # ─── SAVI: Soil Adjusted Vegetation Index ───
        L = 0.5   # Soil brightness correction factor
        savi = ((bands['nir'] - bands['red']) / (bands['nir'] + bands['red'] + L)) * (1 + L)

        # ─── NBR: Normalized Burn Ratio ───
        nbr = (bands['nir'] - bands['swir2']) / (bands['nir'] + bands['swir2'] + eps)

        indices = {'ndvi': ndvi, 'ndwi': ndwi, 'ndsi': ndsi,
                   'evi': evi, 'savi': savi, 'nbr': nbr}

        # Write all indices to a multi-band output
        n_indices = len(indices)
        profile = src.profile.copy()
        profile.update(count=n_indices, dtype='float32', nodata=-9999)

        with rasterio.open(output_path, 'w', **profile) as dst:
            for i, (name, arr) in enumerate(indices.items(), 1):
                dst.write(arr.filled(-9999), i)
                dst.set_band_description(i, name.upper())

    return {name: arr for name, arr in indices.items()}

# indices = compute_indices('sentinel2_scene.tif', 'indices.tif')
```

## Clipping and Masking with Vector Data

```python
import rasterio
from rasterio.mask import mask
import geopandas as gpd
import numpy as np
from shapely.geometry import mapping

def clip_raster_to_polygon(raster_path: str,
                           vector_path: str,
                           output_path: str,
                           all_touched: bool = False):
    """
    Clip a raster to the bounding geometry of a vector file.
    Pixels outside the polygon are set to nodata.
    """
    # Load vector geometries
    gdf = gpd.read_file(vector_path)

    with rasterio.open(raster_path) as src:
        # Ensure vector and raster share the same CRS
        if gdf.crs != src.crs:
            gdf = gdf.to_crs(src.crs)

        # Convert geometries to the format rasterio expects
        geometries = [mapping(geom) for geom in gdf.geometry if geom is not None]

        # Clip: returns (array, transform) for the cropped extent
        out_image, out_transform = mask(
            src,
            geometries,
            crop=True,                  # Shrink extent to bounding box of geometries
            all_touched=all_touched,    # True: include pixels that touch the polygon edge
            nodata=src.nodata if src.nodata else -9999
        )

        # Write clipped raster
        profile = src.profile.copy()
        profile.update(
            height=out_image.shape[1],
            width=out_image.shape[2],
            transform=out_transform
        )
        if src.nodata is None:
            profile.update(nodata=-9999)

        with rasterio.open(output_path, 'w', **profile) as dst:
            dst.write(out_image)

    return out_image, out_transform

# clip_raster_to_polygon('satellite.tif', 'study_area.shp', 'clipped.tif')

# ─── Zonal Statistics: mean/std/count per polygon ───
def zonal_stats(raster_path: str, vector_path: str, band: int = 1) -> gpd.GeoDataFrame:
    """Compute statistics of raster values within each polygon."""
    gdf = gpd.read_file(vector_path).copy()

    with rasterio.open(raster_path) as src:
        if gdf.crs != src.crs:
            gdf = gdf.to_crs(src.crs)

        stats = []
        for idx, row in gdf.iterrows():
            geom = [mapping(row.geometry)]
            try:
                out_image, _ = mask(src, geom, crop=True, all_touched=True)
                values = out_image[band - 1]
                nodata = src.nodata if src.nodata else -9999
                valid  = values[values != nodata]

                stats.append({
                    'count': len(valid),
                    'mean':  float(np.mean(valid)) if len(valid) > 0 else np.nan,
                    'std':   float(np.std(valid))  if len(valid) > 0 else np.nan,
                    'min':   float(np.min(valid))  if len(valid) > 0 else np.nan,
                    'max':   float(np.max(valid))  if len(valid) > 0 else np.nan,
                    'sum':   float(np.sum(valid))  if len(valid) > 0 else np.nan,
                })
            except Exception:
                stats.append({'count': 0, 'mean': np.nan, 'std': np.nan,
                              'min': np.nan, 'max': np.nan, 'sum': np.nan})

        stats_df = gpd.pd.DataFrame(stats)
        return gpd.pd.concat([gdf.reset_index(drop=True), stats_df], axis=1)

# stats = zonal_stats('ndvi.tif', 'municipalities.shp')
# print(stats[['name', 'mean', 'std', 'count']])
```

## Resampling

```python
import rasterio
from rasterio.enums import Resampling
from rasterio.warp import reproject, calculate_default_transform
import numpy as np

def resample_to_resolution(input_path: str,
                           output_path: str,
                           target_res: float,
                           resampling: Resampling = Resampling.bilinear):
    """
    Resample a raster to a specific pixel resolution.
    target_res: desired pixel size in CRS units (meters for UTM, degrees for WGS84)

    Resampling choices:
        Resampling.nearest    → categorical data (land cover class IDs)
        Resampling.bilinear   → continuous data (temperature, elevation)
        Resampling.cubic      → smooth continuous data (best quality, slower)
        Resampling.average    → downsampling continuous data (anti-aliasing)
    """
    with rasterio.open(input_path) as src:
        # Calculate new dimensions
        scale_x = src.res[0] / target_res
        scale_y = src.res[1] / target_res
        new_width  = int(src.width  * scale_x)
        new_height = int(src.height * scale_y)

        # New transform at target resolution
        new_transform = rasterio.Affine(
            target_res, 0, src.transform.c,
            0, -target_res, src.transform.f
        )

        profile = src.profile.copy()
        profile.update(width=new_width, height=new_height, transform=new_transform)

        with rasterio.open(output_path, 'w', **profile) as dst:
            for band_idx in range(1, src.count + 1):
                reproject(
                    source=rasterio.band(src, band_idx),
                    destination=rasterio.band(dst, band_idx),
                    src_transform=src.transform,
                    src_crs=src.crs,
                    dst_transform=new_transform,
                    dst_crs=src.crs,
                    resampling=resampling
                )

# resample_to_resolution('10m_sentinel.tif', '30m_resampled.tif', target_res=30)


# ─── Resample to match another raster (align grids) ───
def resample_to_match(source_path: str, reference_path: str, output_path: str):
    """Resample source raster to exactly match reference raster's grid."""
    with rasterio.open(reference_path) as ref:
        target_transform = ref.transform
        target_crs       = ref.crs
        target_width     = ref.width
        target_height    = ref.height

    with rasterio.open(source_path) as src:
        profile = src.profile.copy()
        profile.update(width=target_width, height=target_height,
                       transform=target_transform, crs=target_crs)

        with rasterio.open(output_path, 'w', **profile) as dst:
            for band_idx in range(1, src.count + 1):
                reproject(
                    source=rasterio.band(src, band_idx),
                    destination=rasterio.band(dst, band_idx),
                    src_transform=src.transform,
                    src_crs=src.crs,
                    dst_transform=target_transform,
                    dst_crs=target_crs,
                    resampling=Resampling.bilinear
                )

# resample_to_match('population.tif', 'landcover.tif', 'population_aligned.tif')
```

## Windowed I/O — Memory-Efficient Processing

```python
import rasterio
from rasterio.windows import Window
import numpy as np

def process_large_raster(input_path: str, output_path: str,
                         block_size: int = 256):
    """
    Process a large raster tile-by-tile without loading into memory.
    Reads in blocks, applies a function, writes results.
    """
    with rasterio.open(input_path) as src:
        profile = src.profile.copy()
        profile.update(dtype='float32')

        with rasterio.open(output_path, 'w', **profile) as dst:
            # Iterate over block windows
            for ji, window in src.block_windows(1):
                # Read this block (all bands)
                data = src.read(window=window).astype(np.float32)

                # ── YOUR COMPUTATION HERE ──
                # Example: normalize each band to [0, 1]
                for b in range(data.shape[0]):
                    band = data[b]
                    valid = band[band != src.nodata] if src.nodata else band
                    if len(valid) > 0:
                        vmin, vmax = valid.min(), valid.max()
                        if vmax > vmin:
                            data[b] = (band - vmin) / (vmax - vmin)
                # ── END COMPUTATION ──

                dst.write(data, window=window)

            print(f"Processed {src.width}x{src.height} in {block_size}x{block_size} blocks")

# process_large_raster('huge_satellite.tif', 'normalized.tif')


# ─── Read specific geographic region without loading full file ───
def read_by_bounds(raster_path: str, bounds: tuple) -> tuple[np.ndarray, dict]:
    """
    Read only the pixels within geographic bounds.
    bounds: (west, south, east, north) in raster's CRS.
    Returns: (array, metadata_dict)
    """
    west, south, east, north = bounds

    with rasterio.open(raster_path) as src:
        # Convert geographic bounds to pixel window
        row_min, col_min = src.index(west, north)   # top-left pixel
        row_max, col_max = src.index(east, south)   # bottom-right pixel

        # Clamp to raster extent
        row_min = max(0, row_min)
        col_min = max(0, col_min)
        row_max = min(src.height, row_max)
        col_max = min(src.width, col_max)

        window = Window(col_min, row_min, col_max - col_min, row_max - row_min)
        data   = src.read(window=window)

        # Transform for this window
        win_transform = src.window_transform(window)

        return data, {
            'transform': win_transform,
            'crs':       src.crs,
            'nodata':    src.nodata,
            'window':    window
        }

# data, meta = read_by_bounds('europe.tif', (10.0, 47.0, 12.0, 49.0))
```

## Visualization

```python
import rasterio
import numpy as np
import matplotlib.pyplot as plt
from matplotlib.colors import Normalize

def plot_raster(raster_path: str, band: int = 1,
                cmap: str = 'viridis', title: str = None):
    """Plot a single band with geographic axes and colorbar."""
    with rasterio.open(raster_path) as src:
        data = src.read(band, masked=True).astype(np.float32)
        bounds = src.bounds

    fig, ax = plt.subplots(figsize=(10, 8))
    im = ax.imshow(
        data,
        extent=[bounds.left, bounds.right, bounds.bottom, bounds.top],
        origin='upper',
        cmap=cmap,
        aspect='equal'
    )
    plt.colorbar(im, ax=ax, label='Value')
    ax.set_xlabel('Longitude' if 'longlat' in str(src.crs) else 'X (m)')
    ax.set_ylabel('Latitude'  if 'longlat' in str(src.crs) else 'Y (m)')
    ax.set_title(title or f'Band {band}')
    plt.tight_layout()
    plt.show()


def plot_rgb(raster_path: str,
             r_band: int = 1, g_band: int = 2, b_band: int = 3,
             stretch: str = 'percentile'):
    """
    Plot a true-color (RGB) composite.
    stretch: 'percentile' (robust, handles outliers) or 'minmax' or None
    """
    with rasterio.open(raster_path) as src:
        r = src.read(r_band).astype(np.float32)
        g = src.read(g_band).astype(np.float32)
        b = src.read(b_band).astype(np.float32)
        bounds = src.bounds

    rgb = np.stack([r, g, b], axis=-1)   # Shape: (H, W, 3)

    if stretch == 'percentile':
        # Per-band percentile stretch — standard for satellite imagery display
        for i in range(3):
            p2, p98 = np.percentile(rgb[:, :, i][rgb[:, :, i] > 0], (2, 98))
            rgb[:, :, i] = np.clip((rgb[:, :, i] - p2) / (p98 - p2 + 1e-10), 0, 1)
    elif stretch == 'minmax':
        for i in range(3):
            vmin, vmax = rgb[:, :, i].min(), rgb[:, :, i].max()
            rgb[:, :, i] = (rgb[:, :, i] - vmin) / (vmax - vmin + 1e-10)

    fig, ax = plt.subplots(figsize=(10, 8))
    ax.imshow(rgb, extent=[bounds.left, bounds.right, bounds.bottom, bounds.top],
              origin='upper', aspect='equal')
    ax.set_title('RGB Composite')
    plt.tight_layout()
    plt.show()

# plot_raster('ndvi.tif', cmap='RdYlGn', title='NDVI')
# plot_rgb('sentinel2.tif', r_band=4, g_band=3, b_band=2)
```

## Practical Workflows

### 1. Land Cover Change Detection

```python
import rasterio
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

def change_detection(path_t1: str, path_t2: str, band: int = 1) -> dict:
    """
    Detect change between two time periods.
    Both rasters must share the same CRS, resolution, and extent.
    Returns change map and statistics.
    """
    with rasterio.open(path_t1) as src1:
        data_t1  = src1.read(band, masked=True).astype(np.float32)
        nodata   = src1.nodata
        profile  = src1.profile.copy()

    with rasterio.open(path_t2) as src2:
        data_t2  = src2.read(band, masked=True).astype(np.float32)

    # Difference map
    change = data_t2 - data_t1

    # Valid pixels only (both periods have data)
    valid  = ~(data_t1.mask | data_t2.mask) if hasattr(data_t1, 'mask') else np.ones_like(change, dtype=bool)
    change_valid = change[valid]

    # Thresholded change classes
    threshold = 0.1   # Adjust based on index scale
    change_class = np.zeros_like(change, dtype=np.int8)
    change_class[change > threshold]   = 1   # Increase (e.g., more vegetation)
    change_class[change < -threshold]  = -1  # Decrease
    # 0 = no significant change

    # Statistics
    stats = {
        'mean_change':     float(change_valid.mean()),
        'std_change':      float(change_valid.std()),
        'pct_increase':    float((change_class[valid] == 1).mean() * 100),
        'pct_decrease':    float((change_class[valid] == -1).mean() * 100),
        'pct_stable':      float((change_class[valid] == 0).mean() * 100),
        'max_increase':    float(change_valid.max()),
        'max_decrease':    float(change_valid.min()),
    }

    # Plot
    fig, axes = plt.subplots(1, 3, figsize=(16, 5))
    axes[0].imshow(data_t1,     cmap='RdYlGn'); axes[0].set_title('T1')
    axes[1].imshow(data_t2,     cmap='RdYlGn'); axes[1].set_title('T2')
    axes[2].imshow(change_class, cmap='RdBu',  vmin=-1, vmax=1)
    axes[2].set_title('Change: Blue=decrease, Red=increase')
    for ax in axes: ax.axis('off')
    plt.tight_layout(); plt.show()

    return stats, change, change_class

# stats, change_map, change_classes = change_detection('ndvi_2020.tif', 'ndvi_2023.tif')
# print(pd.DataFrame([stats]).T)
```

### 2. DEM Analysis — Terrain Derivatives

```python
import rasterio
import numpy as np
from scipy.ndimage import uniform_filter

def dem_analysis(dem_path: str, output_prefix: str = 'terrain'):
    """
    Compute terrain derivatives from a Digital Elevation Model:
    → Slope, Aspect, Hillshade, Curvature
    """
    with rasterio.open(dem_path) as src:
        elev     = src.read(1, masked=True).astype(np.float64)
        res_x, res_y = src.res
        profile  = src.profile.copy()
        profile.update(dtype='float32', count=1)

    # ─── Sobel gradients (dz/dx, dz/dy) ───
    # Using central differences
    dz_dx = np.zeros_like(elev)
    dz_dy = np.zeros_like(elev)
    dz_dx[:, 1:-1] = (elev[:, 2:] - elev[:, :-2]) / (2 * res_x)
    dz_dy[1:-1, :] = (elev[2:, :] - elev[:-2, :]) / (2 * res_y)

    # ─── SLOPE: angle of steepest descent (degrees) ───
    slope_rad = np.arctan(np.sqrt(dz_dx**2 + dz_dy**2))
    slope_deg = np.degrees(slope_rad).astype(np.float32)

    # ─── ASPECT: direction of steepest descent (degrees, 0=North, clockwise) ───
    aspect_rad = np.arctan2(-dz_dy, dz_dx)
    aspect_deg = np.degrees(aspect_rad)
    aspect_deg = (90 - aspect_deg) % 360   # Convert to compass bearing
    aspect_deg = aspect_deg.astype(np.float32)

    # ─── HILLSHADE: simulated illumination ───
    # Standard parameters: sun azimuth=315°, altitude=45°
    az_rad   = np.radians(315)
    alt_rad  = np.radians(45)
    hillshade = (
        np.sin(alt_rad) * np.cos(slope_rad) +
        np.cos(alt_rad) * np.sin(slope_rad) *
        np.cos(az_rad - np.radians(aspect_deg))
    )
    hillshade = np.clip(hillshade * 255, 0, 255).astype(np.float32)

    # ─── Write outputs ───
    outputs = {
        'slope':     slope_deg,
        'aspect':    aspect_deg,
        'hillshade': hillshade,
    }

    for name, arr in outputs.items():
        with rasterio.open(f'{output_prefix}_{name}.tif', 'w', **profile) as dst:
            dst.write(arr, 1)

    # Plot all derivatives
    fig, axes = plt.subplots(2, 2, figsize=(12, 10))
    cmaps = ['gray', 'RdYlGn', 'hsv', 'gray']
    titles = ['Elevation', 'Slope (°)', 'Aspect (°)', 'Hillshade']
    arrays = [elev, slope_deg, aspect_deg, hillshade]

    for ax, arr, cmap, title in zip(axes.flat, arrays, cmaps, titles):
        im = ax.imshow(arr, cmap=cmap)
        plt.colorbar(im, ax=ax, shrink=0.8)
        ax.set_title(title)
        ax.axis('off')
    plt.tight_layout(); plt.show()

    return outputs

# outputs = dem_analysis('dem_alps.tif', output_prefix='alps_terrain')
```

### 3. Full Remote Sensing Pipeline

```python
import rasterio
from rasterio.mask import mask
from rasterio.warp import reproject, calculate_default_transform, Resampling
import geopandas as gpd
import numpy as np
import pandas as pd
from shapely.geometry import mapping

def remote_sensing_pipeline(imagery_path: str,
                            roi_path: str,
                            band_map: dict,
                            target_crs: str = None) -> pd.DataFrame:
    """
    Full pipeline: load → align CRS → clip to ROI → compute indices → zonal stats.

    imagery_path: multispectral satellite image
    roi_path:     vector file with study area polygons (must have 'name' column)
    band_map:     {'red': 3, 'green': 2, 'nir': 4, ...}
    target_crs:   reproject to this CRS (None = use imagery CRS)
    """
    import tempfile, os

    # 1. Load ROI
    roi = gpd.read_file(roi_path)

    # 2. Open imagery and align CRS
    with rasterio.open(imagery_path) as src:
        img_crs = src.crs
        if target_crs:
            from rasterio.crs import CRS
            target = CRS.from_user_input(target_crs)
        else:
            target = img_crs

        roi = roi.to_crs(target)

        # 3. Reproject imagery if needed
        if target != img_crs:
            transform, width, height = calculate_default_transform(
                img_crs, target, src.width, src.height, *src.bounds)
            profile = src.profile.copy()
            profile.update(crs=target, transform=transform, width=width, height=height)

            reprojected_path = tempfile.mktemp(suffix='.tif')
            with rasterio.open(reprojected_path, 'w', **profile) as dst:
                for b in range(1, src.count + 1):
                    reproject(rasterio.band(src, b), rasterio.band(dst, b),
                              src_transform=src.transform, src_crs=img_crs,
                              dst_transform=transform, dst_crs=target,
                              resampling=Resampling.bilinear)
            src_path = reprojected_path
        else:
            src_path = imagery_path

    # 4. Compute indices and zonal stats per polygon
    results = []
    with rasterio.open(src_path) as src:
        eps = 1e-10
        for _, row in roi.iterrows():
            geom = [mapping(row.geometry)]
            try:
                out_img, _ = mask(src, geom, crop=True, all_touched=True)
                # Read bands as float
                bands = {}
                for name, idx in band_map.items():
                    b = out_img[idx - 1].astype(np.float32)
                    nodata = src.nodata if src.nodata else -9999
                    b[b == nodata] = np.nan
                    bands[name] = b

                # Compute NDVI
                if 'nir' in bands and 'red' in bands:
                    ndvi = (bands['nir'] - bands['red']) / (bands['nir'] + bands['red'] + eps)
                    valid = ndvi[~np.isnan(ndvi)]
                    results.append({
                        'name':        row.get('name', 'unknown'),
                        'ndvi_mean':   float(np.nanmean(valid)),
                        'ndvi_std':    float(np.nanstd(valid)),
                        'ndvi_min':    float(np.nanmin(valid)),
                        'ndvi_max':    float(np.nanmax(valid)),
                        'pixel_count': int(len(valid)),
                    })
            except Exception as e:
                results.append({'name': row.get('name', 'unknown'), 'error': str(e)})

    # Cleanup temp file
    if target_crs and os.path.exists(src_path):
        os.remove(src_path)

    return pd.DataFrame(results)

# df = remote_sensing_pipeline(
#     'sentinel2_scene.tif',
#     'municipalities.shp',
#     band_map={'blue': 2, 'green': 3, 'red': 4, 'nir': 8},
#     target_crs='EPSG:32633'
# )
# print(df.sort_values('ndvi_mean', ascending=False))
```

### 4. Raster-to-Vector and Vector-to-Raster

```python
import rasterio
from rasterio.features import shapes, rasterize
import numpy as np
import geopandas as gpd
from shapely.geometry import shape

# ─── RASTER → VECTOR: Extract polygons from classified raster ───
def raster_to_polygons(raster_path: str, band: int = 1) -> gpd.GeoDataFrame:
    """Convert contiguous regions of same value into polygons."""
    with rasterio.open(raster_path) as src:
        data    = src.read(band).astype(np.int32)
        nodata  = int(src.nodata) if src.nodata else -9999
        transform = src.transform
        crs     = src.crs

    # Extract shapes: yields (geometry, value) for each contiguous region
    mask_valid = data != nodata
    polygons = []
    for geom, value in shapes(data, mask=mask_valid, transform=transform):
        polygons.append({
            'geometry': shape(geom),
            'class_id': int(value)
        })

    return gpd.GeoDataFrame(polygons, crs=crs)

# polygons = raster_to_polygons('landcover_classified.tif')
# polygons.to_file('landcover_polygons.shp')


# ─── VECTOR → RASTER: Burn vector attributes into a raster grid ───
def vector_to_raster(vector_path: str,
                     output_path: str,
                     attribute: str,
                     reference_raster: str,
                     fill: float = 0) -> None:
    """
    Rasterize a vector layer onto the grid of a reference raster.
    Each pixel gets the value of the 'attribute' column from the vector.
    """
    gdf = gpd.read_file(vector_path)

    with rasterio.open(reference_raster) as ref:
        gdf = gdf.to_crs(ref.crs)
        transform = ref.transform
        width, height = ref.width, ref.height
        crs = ref.crs

    # Build (geometry, value) pairs for rasterize
    geom_value_pairs = [
        (mapping(row.geometry), row[attribute])
        for _, row in gdf.iterrows()
        if row.geometry is not None
    ]

    # Rasterize
    burned = rasterize(
        geom_value_pairs,
        out_shape=(height, width),
        transform=transform,
        fill=fill,
        dtype=np.float32,
        all_touched=False
    )

    # Write
    with rasterio.open(
        output_path, 'w', driver='GTiff',
        height=height, width=width, count=1,
        dtype=burned.dtype, crs=crs, transform=transform, nodata=fill
    ) as dst:
        dst.write(burned, 1)

# vector_to_raster('population.shp', 'pop_raster.tif',
#                  attribute='pop_density', reference_raster='dem.tif')
```

## Common Pitfalls and Solutions

### Y-Axis Flipped in Plots

```python
import rasterio
import matplotlib.pyplot as plt

# ❌ PROBLEM: imshow default shows row 0 at top (correct for images,
#    but confusing if you also show geographic coordinates)
with rasterio.open('dem.tif') as src:
    data = src.read(1)
    plt.imshow(data)   # Row 0 at top — matches raste

…(truncated)