End-to-end geospatial AI pipeline — from raw satellite bands to land-cover maps, water body detection & change analysis
Ayush Kumar Singh
This pipeline processes multi-spectral satellite imagery from ISRO's ResourceSat-2 LISS-III sensor and classifies every pixel into land-cover types, detects water bodies, and flags areas of change between two dates — all using AI.
It ships in two versions that use different model architectures, explained in full detail below.
Version 1 · run_pipeline.py |
Version 2 · run_pipeline_dl.py |
|
|---|---|---|
| Nickname | Baseline | Paper-Compliant |
| Models | Random Forest · MLP | CNN · LSTM · Transformer · Ensemble |
| Framework | scikit-learn | PyTorch 2.x |
| Water Accuracy | 1.0000 | 1.0000 |
| Land-Cover Accuracy | 0.6694 | 0.6681 |
| Training Time | ~30 sec | ~5–10 min |
| GPU Required | ❌ | ❌ (CPU mode) |
| Matches paper architecture | Partial | ✅ Full |
geodata-processing-ai/
│
├── 🚀 run_pipeline.py ← Version 1 entry point (RF + MLP)
├── 🧠 run_pipeline_dl.py ← Version 2 entry point (CNN + LSTM + Transformer)
│
├── src/
│ ├── config.py ← All paths, thresholds, hyperparameters
│ ├── module_runner.py ← Module loader utility
│ ├── 01_data_ingestion.py ← Load satellite TIFs + CSV datasets
│ ├── 02_preprocessing.py ← Radiometric correction · indices · QA
│ ├── 03_bias_harmonization.py ← Bias analysis · oversampling · scaling
│ ├── 04_ai_models.py ← Version 1: Random Forest + MLP
│ ├── 04_ai_models_dl.py ← Version 2: CNN + LSTM + Transformer
│ ├── 05_change_detection.py ← Multi-temporal change detection
│ ├── 06_visualization.py ← 12 publication-quality figures
│ └── dl_models.py ← PyTorch model definitions
│
├── data/
│ ├── sample/
│ │ ├── water_train.csv ← 1,000 rows · Band2–5 + Water label
│ │ └── vizag_sample_data.csv ← 8,000 rows · 6 bands + 8-class Landcover
│ └── README_data.md ← How to download the TIF files
│
├── tests/ ← 32 unit tests (pytest)
├── notebooks/
│ └── pipeline_walkthrough.ipynb
├── docs/
│ ├── architecture.md
│ ├── modules.md
│ └── results.md
├── assets/ ← Output figure previews
├── MODEL_JUSTIFICATION.txt ← Full explanation of why two versions exist
├── requirements.txt
├── setup.py
├── CHANGELOG.md
└── LICENSE
git clone https://github.com/Ayush-2703/Geodata_Processing.git
cd Geodata_Processing
pip install -r requirements.txtVersion 2 needs PyTorch:
pip install torch
# Haridwar TIF bands
git clone https://github.com/iirs-isro/IIRS_ISRO_Geoprocessing-using-Python.git
cp IIRS_ISRO_Geoprocessing-using-Python/haridwar/haridwarBand*.tif data/
# Water Map reference
git clone https://github.com/iirs-isro/ISRO-Geodata-Processing-using-Python-and-Machine-Learning.git
cp ISRO-Geodata-Processing-using-Python-and-Machine-Learning/Water_Map.tif data/# Version 1 — RF + MLP (~50 seconds, any laptop)
python run_pipeline.py
# Version 2 — CNN + LSTM + Transformer (~10 min, CPU)
python run_pipeline_dl.pyBoth versions share the same 6-module structure. Only Module 4 differs.
[ Satellite TIFs ] [ Water Map ] [ water_train.csv ] [ vizag.csv ]
└──────────────────┴────────────────┴──────────────────┘
│
┌───────────────▼────────────────┐
│ MODULE 1 · Data Ingestion │
│ Raster loader · CSV loader │
└───────────────┬────────────────┘
│
┌───────────────▼───────────────────────┐
│ MODULE 2 · Pre-processing & QA │
│ DN→Reflectance · Gaussian filter │
│ Cloud/shadow detection (>92% acc) │
│ NDVI · NDWI · MNDWI · NDBI · NBR │
│ QA confidence raster [0–1] │
└───────────────┬───────────────────────┘
│
┌───────────────▼───────────────────────┐
│ MODULE 3 · Bias Mitigation │
│ CV bias check · Oversampling │
│ StandardScaler · Feature matrix │
└───────────────┬───────────────────────┘
│
┌──────────────────────┴──────────────────────┐
│ │
┌────────────▼──────────────┐ ┌───────────────────▼──────────────────┐
│ MODULE 4 · Version 1 │ │ MODULE 4 · Version 2 │
│ │ │ │
│ RandomForestClassifier │ │ SpectralCNN (U-Net encoder-dec) │
│ n_estimators = 200 │ │ TemporalLSTM (BiLSTM + GRU) │
│ max_depth = 15 │ │ SpectralTransformer (MHA, Pre-LN) │
│ │ │ EnsembleFusion (learnable weights) │
│ MLPClassifier │ │ GradientSaliency (XAI) │
│ layers: 128→64→32 │ │ │
│ early_stopping = True │ │ PyTorch 2.x · ~10 min · CPU only │
│ │ └───────────────────┬──────────────────┘
│ ~30 sec · no GPU │ │
└────────────┬──────────────┘ │
└──────────────────────┬──────────────────────┘
│
┌───────────────▼───────────────────────┐
│ MODULE 5 · Change Detection │
│ Δ-Index differencing (T2 − T1) │
│ Vegetation loss · Urban expansion │
│ Water change · Anomaly detection │
│ Morphological cleaning │
└───────────────┬───────────────────────┘
│
┌───────────────▼───────────────────────┐
│ MODULE 6 · Visualisation │
│ 12 publication-quality figures │
│ GeoTIFF prediction outputs │
└───────────────────────────────────────┘
Fast, reliable, and highly accurate on the available ISRO tabular datasets. The right tool for the available data format.
# Water body classification
RandomForestClassifier(n_estimators=200, max_depth=15, class_weight="balanced")
# Land-cover (8-class)
MLPClassifier(hidden_layer_sizes=(128, 64, 32), max_iter=500, early_stopping=True)Why this version exists: The ISRO training CSVs are tabular — one row = one pixel's band values. CNNs require 2D spatial patches; LSTMs require time-series with multiple dates. RF and MLP are the scientifically correct choice for this data format and run in ~50 seconds on any laptop.
Full technical reasoning →
MODEL_JUSTIFICATION.txt
Implements the exact architecture stated in the research paper using PyTorch 2.x.
Paper: "U-Net and ResNet architectures to extract spatial features at various scales"
Input (B, 4, H, W)
Encoder: [Conv→BN→ReLU] × 3 + MaxPool
↕ skip connections
Decoder: [ConvTranspose] × 3 + skip concat
Output (B, num_classes, H, W)
Paper: "LSTM and GRU variants to capture temporal dynamics in time-series data"
Input (B, T, features)
BiLSTM × 2 layers hidden=128 dropout=0.3
GRU refinement
LayerNorm → Linear(128→64→classes)
Output (B, num_classes)
Paper: "Self-attention mechanisms facilitate integration of contextual information across spatial and temporal dimensions"
Input (B, seq_len, 1) ← each band = one token
Linear projection → d_model=64
Positional encoding (sinusoidal)
TransformerEncoder × 2 nhead=4 FFN=256 GELU Pre-LN
Global average pool
Linear → num_classes
Output (B, num_classes)
Paper: "multi-modal fusion framework that achieves both computational efficiency and analytical strength"
CNN probs ──┐
LSTM probs ──┼──► Learnable weighted softmax ──► Final prediction
TRF probs ──┘ (weights optimised via AdamW on training set)
Paper: "explainable AI components mitigate the black box issues linked with deep learning"
saliency = GradientSaliency(model).compute(x, target_class=1)
# Returns same shape as input — high values = most important features| Model | Version | Accuracy | F1-Macro | F1-Weighted |
|---|---|---|---|---|
| Random Forest | v1 | 1.0000 | 1.0000 | 1.0000 |
| MLP | v1 | 1.0000 | 1.0000 | 1.0000 |
| CNN | v2 | 1.0000 | 1.0000 | 1.0000 |
| LSTM | v2 | 1.0000 | 1.0000 | 1.0000 |
| Transformer | v2 | 1.0000 | 1.0000 | 1.0000 |
| Ensemble | v2 | 1.0000 | 1.0000 | 1.0000 |
| Model | Version | Accuracy | F1-Macro |
|---|---|---|---|
| Random Forest | v1 | 0.6544 | 0.6525 |
| MLP | v1 | 0.6694 | 0.6681 |
| CNN | v2 | 0.6681 | 0.6670 |
| LSTM | v2 | in training | — |
| Transformer | v2 | in training | — |
Classes:
Cropland · Forest · Grassland · Shrubland · Urban · Bare Land · Water · Wetland
| Metric | Value |
|---|---|
| Scene | 1,151 × 1,151 px · EPSG:4326 |
| Total changed | 78,825 px (5.95%) |
| Vegetation loss | 40,398 px · 51.1% |
| Urban expansion | 39,222 px · 49.7% |
| Water change | 348 px · 0.4% |
| QA confidence | 1.000 (no cloud) |
| Training | Full pipeline | |
|---|---|---|
| v1 (RF + MLP) | ~30 sec | ~50 sec |
| v2 (DL models) | ~5–10 min | ~12 min |
| File | Type | Rows / Size | Description |
|---|---|---|---|
haridwarBand2–5.tif |
GeoTIFF | 1151×1151 px | ResourceSat-2 LISS-III · 04-May-2019 |
Water_Map.tif |
GeoTIFF | 1151×1151 px | Binary reference water mask |
water_train.csv |
CSV | 1,000 rows | Band2–5 + Water label (50/50) |
vizag_sample_data.csv |
CSV | 8,000 rows | 6 bands + 8-class Landcover |
| Index | Formula | Application |
|---|---|---|
| NDVI | (NIR−Red)/(NIR+Red) | Vegetation density |
| NDWI | (Green−NIR)/(Green+NIR) | Surface water |
| MNDWI | (Green−SWIR)/(Green+SWIR) | Modified water |
| NDBI | (SWIR−NIR)/(SWIR+NIR) | Built-up / urban |
| NBR | (NIR−SWIR)/(NIR+SWIR) | Burn / bare soil |
# Both runners accept the same flags:
--tif NAME=PATH Add extra satellite band
--csv water=PATH Override water training CSV
--csv vizag=PATH Override land-cover CSV
--before-tif NAME=PATH Real T1 bands for two-date change detection
--no-viz Skip figure generationExamples:
# Fastest possible run
python run_pipeline.py --no-viz
# Real two-date change detection
python run_pipeline_dl.py \
--before-tif B2_Green=/T1/b2.tif B3_Red=/T1/b3.tif \
B4_NIR=/T1/b4.tif B5_SWIR=/T1/b5.tif
# Use your own data
python run_pipeline.py \
--csv water=/my/water.csv \
--tif B2_Green=/my/b2.tif B3_Red=/my/b3.tifpip install -r requirements-dev.txt
pytest tests/ -v| File | Tests | Covers |
|---|---|---|
test_ingestion.py |
6 | CSV loading, TIF reading, label validation |
test_preprocessing.py |
9 | Index ranges, cloud detection, normalisation |
test_models.py |
10 | RF accuracy, MLP convergence, feature importance |
test_change_detection.py |
7 | Delta computation, thresholds, anomaly detection |
32 / 32 passing ✅
# Both versions
rasterio>=1.3.0 scikit-learn>=1.2.0 pandas>=1.5.0
numpy>=1.23.0 matplotlib>=3.6.0 scipy>=1.9.0
# Version 2 only
torch>=2.0.0
- Ms. Garima Srivastava — guidance and supervision throughout
- IIRS / ISRO — Haridwar satellite imagery and online course datasets
- Open-source community — scikit-learn, PyTorch, rasterio, NumPy, Matplotlib
Made with ❤️ by Ayush Kumar Singh Amity University Uttar Pradesh · B.Tech Artificial Intelligence · 2025
⭐ Star this repo if it helped you!