🌍 Advanced Land Use Detection & Classification (EuroSAT)

📖 Overview

This project implements an advanced intelligent system for Land Use and Land Cover (LULC) classification using the EuroSAT Sentinel-2 satellite dataset.

The study conducts a rigorous comparative analysis between RGB (3-band) and Multispectral (13-band) imagery across distinct paradigms:

1. Classical Machine Learning: SVM, Random Forest, KNN.
2. Deep Learning (CNNs): Custom Simple CNN, ResNet50 (Transfer Learning), and ConvNeXt.
3. Transformers: Vision Transformer (ViT) feature extraction.

Key Achievement: The system achieved a peak accuracy of 97.00% using Multispectral ResNet50 and demonstrated that ConvNeXt offers superior robustness to sensor degradation (blur/low resolution).

---

📑 Table of Contents

• Dataset
• Methodology
• Installation
• Usage
• Results & Performance
• Robustness Analysis
• Visualizations (EDA)
• Project Structure
• Future Work

---

🛰 Dataset

The project utilizes the EuroSAT dataset based on Sentinel-2 satellite imagery.

• Total Images: 27,000
• Classes (10): Annual Crop, Forest, Herbaceous Vegetation, Highway, Industrial, Pasture, Permanent Crop, Residential, River, Sea/Lake.
• Spatial Resolution: 10m - 60m per pixel.
• Modalities Used:
  • RGB: Standard visible spectrum (Bands 4, 3, 2).
  • Multispectral: Full 13-band spectrum including Near-Infrared (NIR) and Short-Wave Infrared (SWIR).

---

⚙️ Methodology

1. Preprocessing

• RGB: Gaussian Blur ($3\times3$) for noise removal, CLAHE for contrast enhancement, and ImageNet Normalization.
• Multispectral: Band-wise Min-Max scaling to $[0, 1]$.

2. Feature Extraction & Architectures

• Classical: GLCM (Texture), LBP, and Spectral Statistics (Mean/Std/NDVI).
• Simple CNN: A custom 3-layer CNN trained from scratch (Lightweight).
• ResNet50: Transfer learning from ImageNet, adapted for 13-channel input.
• ConvNeXt: Modern CNN architecture using large ($7\times7$) kernels to mimic Vision Transformers.
• ViT (Vision Transformer): Patch-based feature extraction ($16\times16$ patches).

---

💻 Installation

1. Clone the repository:

   git clone [https://github.com/your-username/EuroSAT-Land-Use-Classification.git](https://github.com/your-username/EuroSAT-Land-Use-Classification.git)
   cd EuroSAT-Land-Use-Classification

2. Install dependencies:

   pip install -r requirements.txt

   Key libraries: torch, torchvision, rasterio, scikit-learn, seaborn, albumentations.

---

🚀 Usage

I used Google Colab with T4 GPUs! The jupyter notebooks are ready to use in that environment

---

📊 Results & Performance

We evaluated models based on Accuracy, F1-Score, and Inference Efficiency.

Architecture	Modality	Accuracy	F1-Score	Verdict
ResNet50	Multispectral	97.00%	0.97	🏆 Best Overall
Simple CNN	Multispectral	94.00%	0.93	🚀 Best Efficiency
ConvNeXt	RGB	92.00%	0.92	Strong RGB Baseline
ViT + SVM	RGB	96.00%	0.96	Excellent Feature Extractor
Classical SVM	Multispectral	86.00%	0.86	Good Baseline
KNN	Deep Features	63.00%	0.60	Poor (Curse of Dimensionality)

Key Findings

• The Spectral Leap: Adding NIR/SWIR bands improved accuracy by ~15% for vegetation classes (Forest vs. Pasture).
• Small Models Win: The custom Simple CNN (trained from scratch) achieved 94% on Multispectral data, outperforming complex hybrids like ViT+RF (62%).

---

🛡 Robustness Analysis

We stress-tested models against synthetic degradations to simulate real-world sensor issues.

Degradation Type	ResNet50 (MS)	ConvNeXt (MS)	ViT+SVM (RGB)
Baseline Accuracy	97.26%	92.00%	95.80%
Gaussian Noise	11.19% (Fail)	55.06%	98.50% (Robust)
Cloud Shadows	82.37% (Pass)	65.30%	3.90% (Fail)
Blur (Low Res)	53.89% (Fail)	91.61% (Robust)	28.10% (Fail)

Insight: ConvNeXt is the best choice for low-quality/blurry satellite feeds, while ViT is immune to sensor noise.

---

📈 Visualizations (EDA)

(Placeholders for your generated plots - add these to an /images folder)

• Spectral Signatures: Visual proof that Vegetation (Band 8 spike) is separable from Urban areas.
• t-SNE Clustering: 2D projection showing class separability.
• Confusion Matrices: Detailed error analysis per class.

---

🔮 Future Work

1. SatMAE (Masked Autoencoders): Implementing self-supervised learning to reconstruct masked patches for better physics understanding.
2. Ensembling: Combining ResNet50 (Spatial Expert) and Simple MS-CNN (Spectral Expert) to maximize robustness.
3. Dimensionality Reduction: Applying PCA to Deep Features before passing them to Classical Classifiers (SVM/KNN).

---

📜 License

This project is licensed under the MIT License.

---