Merge branch 'jpramil-review-jp'

Author: micedre

1-acquisition.qmd

@@ -9,13 +9,13 @@ format:
 ::: {.callout-note}
 ## Where this chapter fits
 
-Training a segmentation model needs a dataset of `(image, label)` pairs — satellite tiles on one side, land-cover maps on the other. **This chapter builds that dataset.** By the end, you'll know how to pull a Sentinel-2 tile for any European city, fetch its matching CLC+ label, and confirm the two are perfectly aligned pixel-for-pixel.
+Training a segmentation model needs a dataset of `(image, label)` pairs — satellite tiles on one side, land-cover maps on the other. **This chapter builds that dataset.** By the end, you'll know how to pull a Sentinel-2 tile for any European city, fetch its matching CLC + label (Corinne Land Cover), and confirm the two are perfectly aligned pixel-for-pixel.
 
 ```{mermaid}
 %%| eval: true
 flowchart LR
     C1["Chapter 1Build the(image, label) dataset"]:::current
-    C2["Chapter 2 (optional)Train a modelon the dataset"]
+    C2["Chapter 2 (optional)Train a model on the dataset"]
     C3["Chapter 3Use a trained modelto predict new masks"]
     C4["Chapter 4Compute regionalland-cover statistics"]
     C1 --> C2 --> C3 --> C4
@@ -56,15 +56,11 @@ These properties make Sentinel-2 a reference dataset for land-cover mapping, agr
 ![The Twin-Satellite Sentinel-2 Orbital Configuration (Credits: Astrium GmbH)](https://sentiwiki.copernicus.eu/__attachments/1687033/image-20230605-102800.png?inst-v=7382be46-c982-4c77-bf23-f3ed1f2e446f)
 
 ::: {.callout-tip}
-**Learn more:** [Sentinel-2 Mission information](https://sentiwiki.copernicus.eu/web/s2-mission) · [Sentinel Hub custom scripts gallery](https://custom-scripts.sentinel-hub.com/)
+**Learn more:** [Sentinel-2 Mission information](https://sentiwiki.copernicus.eu/web/s2-mission) (see in particular the [Sentinel-2 Spectral Response Functions (S2-SRF) document](https://sentiwiki.copernicus.eu/__attachments/1692737/COPE-GSEG-EOPG-TN-15-0007%20-%20Sentinel-2%20Spectral%20Response%20Functions%202024%20-%204.0.xlsx?inst-v=29a9468e-c82d-4e65-9e31-d654700543f9) for the spectral bands) · [Sentinel Hub custom scripts gallery](https://custom-scripts.sentinel-hub.com/)
 :::
 
 ### Spectral bands
 
-::: {.callout-tip}
-**Official reference:** [Sentinel-2 Mission information](https://sentiwiki.copernicus.eu/web/s2-mission), specifically [Sentinel-2 Spectral Response Functions (S2-SRF) document](https://sentiwiki.copernicus.eu/__attachments/1692737/COPE-GSEG-EOPG-TN-15-0007%20-%20Sentinel-2%20Spectral%20Response%20Functions%202024%20-%204.0.xlsx?inst-v=29a9468e-c82d-4e65-9e31-d654700543f9)
-:::
-
 A **spectral band** is a narrow window of wavelengths that the sensor records separately. Sentinel-2's MultiSpectral Instrument (MSI) captures **13 bands** spanning visible, near-infrared (NIR) and short-wave infrared (SWIR) light — from 443 nm to 2190 nm. Each band is tuned to a specific physical property of the surface or atmosphere, and the table below groups them by **native spatial resolution**, which reflects a hardware trade-off made by ESA:
 
 - **10 m — the "core" mapping bands.** `B2` (blue), `B3` (green), `B4` (red) and `B8` (NIR). These are the bands you'll use most often to look at the landscape: true-colour RGB composites, NDVI, and any high-detail analysis of vegetation, water or built-up areas.
@@ -74,9 +70,11 @@ A **spectral band** is a narrow window of wavelengths that the sensor records se
 ::: {.callout-note}
 **L1C vs L2A — two processing levels.** Copernicus distributes Sentinel-2 data at two processing levels. **Level-1C** contains top-of-atmosphere reflectance for all 13 bands. **Level-2A** applies atmospheric correction and delivers surface reflectance, which is what we want for land-cover mapping. L2A drops `B10` (the cirrus band — only useful for cloud detection, not for the surface) and delivers **12 spectral bands**. The training dataset in this tutorial is built from L2A products: the pre-processed GeoTIFFs stack these 12 spectral bands and append **two derived vegetation/water indices, NDVI and NDWI**, for a total of **14 channels**.
 
+A **GeoTIFF** is a standard TIFF image file with extra embedded metadata describing *where* the image sits on Earth — coordinate reference system (CRS), pixel size, and an affine transform that maps each `(row, column)` to a geographic `(x, y)` coordinate. In other words: a regular image plus the spatial information needed to overlay it correctly on a map. The format also supports multiple bands in a single file, which is why we can pack the 14 channels into one `.tif`
+
 NDVI (Normalised Difference Vegetation Index) and NDWI (Normalised Difference Water Index) are computed once at preprocessing time so the model can use them as input features without having to re-derive them from the spectral bands at every batch.
 
-*Regenerating from scratch:* `src/download_region.py` follows the same convention and produces 14-channel float64 GeoTIFFs identical in layout to those on `projet-funathon`.
+`src/download_region.py` follows the same convention and produces 14-channel float64 GeoTIFFs identical in layout to those on `projet-funathon`.
 :::
 
 | Band | Central <br> wavelength (nm) | Resolution (m) | Description | Use / Application |
@@ -101,6 +99,8 @@ NDVI (Normalised Difference Vegetation Index) and NDWI (Normalised Difference Wa
 
 ::: {.callout-note}
 **Think of a raster as a spreadsheet with coordinates.** A regular spreadsheet has rows and columns filled with numbers. A raster is the same idea, except each cell is a *pixel* located at a specific place on Earth, and you can stack multiple layers (bands) on top of each other — one layer per spectral band.
+
+![Single-band vs. multi-band raster. *Source: [Earth Lab — Earth Data Science](https://earthdatascience.org/courses/earth-analytics/multispectral-remote-sensing-data/naip-imagery-raster-stacks-in-r/) (CC BY-SA 4.0) / NEON.*](assets/RGB-bands-raster-stack.jpg){width=70% fig-align="center"}
 :::
 
 Key concepts:
@@ -1551,6 +1551,8 @@ m
 
 In supervised machine learning, you need **pairs** of inputs and labels to train a model. The inputs are the Sentinel-2 satellite images; the labels are provided by [CLC+ Backbone](https://land.copernicus.eu/en/products/clc-backbone), a high-resolution land-cover product from the [Copernicus Land Monitoring Service](https://land.copernicus.eu/en).
 
+**What does CLC stand for?** CLC is short for **CORINE Land Cover** (*Coordination of Information on the Environment — Land Cover*), the European reference inventory of land cover and land use. First produced in 1990 and updated roughly every six years, it has long been the standard product for tracking land-cover change across Europe. **CLC+ Backbone** is the next-generation evolution of this programme: instead of the original 44 hierarchical classes mapped at 100 m resolution from manually delineated polygons, it delivers a **pixel-level raster at 10 m resolution** with a simplified set of 10 land-cover classes, derived automatically from Sentinel-2 imagery — which makes it directly suitable as ground truth for a deep-learning segmentation model.
+
 CLC+ Backbone provides **pixel-level** land-cover classification across Europe at **10 m resolution**, making it an ideal ground-truth dataset for training and evaluating segmentation models. The labels are temporally aligned with the satellite imagery (same year), ensuring that the land-cover classification corresponds to what the satellite actually observed.
 
 **Why 2021?** CLC+ Backbone's most recent public release at the time of writing is the **2021** edition, so that's the year we use for labels. Each label is paired with a Sentinel-2 image from the **same year** — land cover actually changes between years (new construction, deforestation, crop rotation), and mismatched pairs would teach the model to predict yesterday's ground truth for today's image. The 2024 Sentinel-2 imagery you saw in the earlier exercises is kept for Chapter 3, where we run the trained model on *new* images that have no label — which is the whole point of training a model in the first place.