Sentinel-2 End-to-End Python Workflow: 47 Indices & 8-Class Land Classification

Introduction

This project builds an end-to-end automated workflow for Sentinel-2 L2A imagery, using Haikou City as a case study. The pipeline covers data acquisition, preprocessing, 47 remote sensing indices calculation, index applications, visualization, and 8-class land classification compliant with GB/T 21010-2017.

Data and Environment

Imagery source: Sentinel-2 L2A (surface reflectance, atmospherically corrected)

Data portal: Copernicus Data Space (OData/OAuth2.0)

Study area: Haikou City, spanning UTM tiles T49QDB and T49QDC

Bands: 12 bands (B1–B12) resampled to 10 m resolution

Classification standard: GB/T 21010-2017 "Current Land Use Classification" three-level system

Runtime: Python 3.10+, Windows/Linux

UI framework: PySide6 (Qt 6) for visualization script

Module Overview

Preprocessing and Indices (9 scripts)

1_sentinel_download.py

– Automated Sentinel-2 query and download (OData filtering, OAuth2.0, streaming download, Excel manifest) 2_resampling_stacking.py – Resample multi-resolution bands to 10 m and stack into 12-band TIFF; process SCL layer 3_cloud_shadow_masking.py – Mask clouds, cloud shadows, cirrus using SCL scene classification (block processing, default mask codes 3/8/9/10) 4_cloud_free_image_mosaicking.py – Multi-temporal cloud-free filling: tile grouping, adaptive base selection, rule-based/adaptive dual strategy, fill quality report 5_tiff_merge.py – Cross-tile mosaicking: automatic boundary calculation, four overlap strategies (nodata_first/first/last/mean), distance-weighted feathering 6_tiff_shp_cut.py – Clip by Shapefile administrative boundary: geopandas filtering, CRS auto-conversion, rasterio.mask 7_remote_sensing_index.py – Batch index calculation entry: calls remote_sensing_index.py to compute 47 indices and output TIFFs 8_remote_sensing_index_application.py – Index applications: single-index classification, multi-index weighted evaluation, national-standard land classification (modules 1/2/3/all) 9_remote_sensing_index_ui.py – PySide6 interactive viewer: three-level tree menu, pyramid zoom, continuous/discrete dual-mode rendering, dynamic legend

Land Classification (4 scripts)

13_water_extract.py

– Water prior extraction: multi-index voting + KMeans clustering + river neighborhood growth + roof suppression 14_road_extract.py – Road prior extraction: directional contrast line detection + trunk/village road morphological growth + building exclusion 15_building_extract.py – Building prior extraction: roof seeds + density threshold + expansion to impervious surfaces + red roof identification 16_land_classification.py – Land classification CLI: reads indices from result/ and priors from application/, outputs three-level classification

Core Libraries

remote_sensing_index.py

– RemoteSensingIndex class: unified engine for 47 indices, reflective invocation (getattr), safe division, nodata masking land_classification.py – LandClassificationApplication / WeightedIndexLandClassifier: candidate rules, Up/Down/Tri physical response scoring, P2–P98 robust standardization, area quotas, mutually exclusive allocation, sub-class subdivision

Installation and Usage

Environment Setup

git clone https://gitee.com/baoqiangwang/remote-sensing-index-application.git
cd remote-sensing-index-application
python -m venv .venv
# Windows: .venv\Scripts\activate
# Linux: source .venv/bin/activate
pip install -r requirements.txt

Note: requirements.txt locks only direct imports (rasterio, numpy, geopandas, pandas, requests, openpyxl, tqdm, scipy, opencv-python, scikit-learn, PySide6); transitive dependencies resolved by pip.

Full Pipeline Commands

# Stage 1: Data acquisition & preprocessing
# 1. Download Sentinel-2 (edit user config at top of script)
python 1_sentinel_download.py

# 2. Resample & stack (modify RAW_DATA_ROOT / OUTPUT_ROOT)
python 2_resampling_stacking.py

# 3. Cloud/shadow masking (modify INPUT_ROOT / OUTPUT_ROOT)
python 3_cloud_shadow_masking.py

# 4. Cloud-free mosaicking (modify MASKED_ROOT / OUTPUT_ROOT / FILL_MAPPING)
python 4_cloud_free_image_mosaicking.py

# 5. Tile merging (optional feathering -f)
python 5_tiff_merge.py T49QDB_202605_masked.tif T49QDC_202605_masked.tif -o T49QDB_QDC_202605.tif -m nodata_first -f

# 6. Shapefile clipping (-x admin code 460100 for Haikou)
python 6_tiff_shp_cut.py -t T49QDB_QDC_202605.tif -s haikou.shp -o T49QDB_QDC_202605_haikou.tif -x 460100

# Stage 2: Indices & applications
# 7. Compute 47 indices (omit -i for all)
python 7_remote_sensing_index.py -t result/T49QDB_QDC_202605_haikou.tif -o result -i NDVI NDWI EVI NDBI MNDWI

# 8. Index applications (module 1=classification 2=evaluation 3=standard all=all)
python 8_remote_sensing_index_application.py -t result/T49QDB_QDC_202605_haikou.tif -o application -i result -m all

# 9. Launch visualization UI
python 9_remote_sensing_index_ui.py

# Stage 3: 8-class land classification
# 13/14/15. Extract water, road, building priors (output to application/ subdirs)
python 13_water_extract.py -t result/T49QDB_QDC_202605_haikou.tif
python 14_road_extract.py -t result/T49QDB_QDC_202605_haikou.tif
python 15_building_extract.py -t result/T49QDB_QDC_202605_haikou.tif

# 16. Land classification: reads indices from result/ + priors from application/, outputs three-level classification
python 16_land_classification.py -t result/T49QDB_QDC_202605_haikou.tif -i result -o application

Remote Sensing Indices and Classification Results

47 Indices Overview

The RemoteSensingIndex class computes 47 indices grouped into six categories:

Basic vegetation ratio (1): SR – quick vegetation/bare soil discrimination

General normalized vegetation (17): NDVI, EVI, NDRE, GNDVI, WDRVI, CIgreen, ARI… – vegetation vigor, chlorophyll, nitrogen

Soil-adjusted vegetation (7): SAVI, MSAVI, OSAVI, PVI, MCARI, TCARI/OSAVI… – low cover/early-stage monitoring

Atmospheric/aerosol corrected (4): ARVI, SARVI, AFRI, VARI – cloudy/hazy conditions

Water/drought/salinity stress (7): MSI, NDMI, LSWI, AWEIsh, AWEInsh, SMMI… – water, drought, salinity assessment

Bare soil/water/built-up classification (11): NDBI, BSI, BI, MNDWI, NDWI, IBI, NBR, MVI… – land cover classification, feature extraction

Formulas are implemented in remote_sensing_index.py; reflectance values are divided by 10000 before calculation, with _safe_divide handling division by zero and _apply_mask flagging invalid pixels.

Eight-Class Land Classification System

Classification follows GB/T 21010-2017 three-level hierarchy: Level 1 (3 classes) → Level 2 (12 classes) → Level 3 (35 valid codes). Allocation order: Transportation → Water → Built-up (priors locked first), then the remaining pixels compete among Cropland, Orchard, Forest, Grassland, Wetland , finally subdivided by official area quotas.

Cropland (3 sub-classes): indices NDVI, EVI, NDRE, LSWI, NDMI; five-class competition, suppresses urban high-density misclassification

Orchard (4 sub-classes): NDVI, EVI, NDRE, NBR, LSWI; high vegetation threshold, emphasizes perennial vegetation

Forest (4 sub-classes): NDVI, EVI, NDRE, NBR; prefers high cover, suppresses bare soil and built-up

Grassland (2 sub-classes): NDVI, EVI, NDRE, BSI, NDMI; peak-shaped vegetation response, avoids confusion with dense forest

Wetland (4 sub-classes): MNDWI, LSWI, NDMI, MVI, NDWI; dual candidate channels for vegetated and open wetland

Water (5 sub-classes): NDWI, MNDWI, AWEIsh, AWEInsh; water prior + multi-index voting + 320 m hydrological neighborhood + roof suppression

Built-up (5 sub-classes): BI, NDBI, BSI, low NDVI; building prior bonus, road penalty, converted at 40% cover ratio

Transportation (6 sub-classes): NDBI, BSI, medium BI, low water indices; road prior bonus, building penalty, water hard exclusion

Known limitations: Single-date imagery lacks phenological information; 10 m mixed pixels hinder narrow feature separation; LSWI duplicates NBR , IBI duplicates BSI (identical formulas); area quotas control quantity but not spatial accuracy; empirical thresholds calibrated for Haikou's tropical coastal scene, requiring recalibration for transfer.

Directory Structure

remote-sensing-index-application/
├── README.md
├── requirements.txt
├── assets/
│   ├── qrcode_mp.png          # WeChat official account QR code
│   └── qrcode_wechat.png      # WeChat QR code
├── 1_sentinel_download.py
├── 2_resampling_stacking.py
├── 3_cloud_shadow_masking.py
├── 4_cloud_free_image_mosaicking.py
├── 5_tiff_merge.py
├── 6_tiff_shp_cut.py
├── 7_remote_sensing_index.py
├── 8_remote_sensing_index_application.py
├── 9_remote_sensing_index_ui.py
├── 13_water_extract.py
├── 14_road_extract.py
├── 15_building_extract.py
├── 16_land_classification.py
├── remote_sensing_index.py
└── land_classification.py

All scripts remain flat in the root directory to ensure cross-script import and importlib loading work correctly.

License

Apache License 2.0.

Contact

The project accompanies two WeChat article series: 【Remote Sensing Image Analysis】 (9 articles) covering the full workflow from download to visualization, and 【One Scene, Eight Classes】 (9 articles) explaining how a single satellite image is split into the eight major land categories. The complete open-source repository, sample data, and debug versions are maintained via the WeChat official account "Python and Big Data Analysis" (reply keyword remotesensingindexapplication).