mahfuj735/LiteRS-Mamba-Remote-Sensing

🛰️ LiteRS-Mamba: Real-Time Multi-Platform Remote Sensing Land Cover Intelligence for Bangladesh

Master's / Academic Thesis Research Repository
Author: Md. Mahfujul Karim Sheikh (Student ID: CSE11220120735)
Project: Multi-Platform Remote Sensing Land Cover Classification in Bangladesh: Benchmarking Deep Vision Paradigms, Proposing LiteRS-Mamba, and Test-Time Spectral Radiometric Adaptation.

---

🌟 Executive Summary

Remote sensing land-use and land-cover (LULC) classification in the dynamic deltaic terrain of Bangladesh is severely impaired by cross-sensor radiometric discrepancies across web-mapping satellite providers (ArcGIS Map, Bing Maps, Google Maps). Conventional deep learning models suffer catastrophic cross-platform domain collapse when evaluated across sensors.

This repository provides LiteRS-Mamba, an ultra-compact $2.05\text{ M}$ parameter architecture engineered specifically for low-power edge UAVs, survey drones, and satellite edge computing:

1. Multi-Scale Stem (MS-Stem): Dual $3\times3$ and $5\times5$ convolutional tokenizer resolving fine-grained riparian boundaries and parcel edges.
2. 4-Way Omnidirectional Selective State Space Engine (O4-SSM): Scans 2D visual tokens along horizontal ($x$), vertical ($y$), and reverse pathways in parallel, with learnable softmax directional fusion.
3. Test-Time Radiometric Adaptation (TTA-RHM): Zero-shot cumulative distribution matching recovering $+45.65%$ accuracy on raw Google Maps imagery without retraining.
4. Token-Level SSM Grad-CAM: Explainable AI framework offering physical topological verification and forensic failure diagnosis.
5. Universal ONNX & INT8 Edge Deployment: Serialized graph verified to $3.34 \times 10^{-6}$ mathematical parity, quantized to $8.58\text{ MB}$ ($23.23%$ storage reduction) with only $0.29%$ accuracy delta ($97.43% \rightarrow 97.14%$) and $99.71%$ decision agreement.

---

🗺️ The 7 Canonical Bangladesh Land Cover Classes

Class Icon	Class Name	Bengali Meaning	Key Ecological & Physical Characteristics
🌾	Agricultural_Land	ফসলি জমি / কৃষি ভূমি	Alluvial crop fields, seasonal boro/aman paddy cultivation, silt floodplains.
🏖️	Beaches_Coastline	সমুদ্র সৈকত ও উপকূলরেখা	Sandy shores, intertidal mudflats in Cox's Bazar, Kuakata & Sundarbans delta.
🏙️	Built_up_Areas	শহরাঞ্চল ও পাকা স্থাপনা	High-density urban dwellings, impervious rooftops, asphalt roads, transport corridors.
🌿	Grasslands_Open_Land	ঘাসজমি ও চরাঞ্চল	Vegetated riverine chars, grazing ground, and riparian bars in Jamuna/Padma basin.
⛰️	Hills_Mountains	পাহাড় ও টিলা অঞ্চল	Tertiary folding ranges, tea gardens, tropical forest slopes in CHT & Sylhet.
🌲	Trees_Forests	গাছপালা ও বনাঞ্চল	Dense mangrove canopies of Sundarbans, Sal forests of Madhupur, social plantations.
🌊	Water_Bodies	নদী, খাল ও জলাশয়	Major braided rivers (Padma, Jamuna, Meghna), haors, beels, and coastal estuaries.

---

📊 Comprehensive Experimental Benchmark Summary

1. In-Domain Classification vs. Deep Vision Baselines (ArcGIS Map)

Model Architecture	Paradigm	Parameters (M)	In-Domain OA (%)	Macro-F1 (%)	Storage Footprint (MB)
ResNet-50	Classical CNN	25.56	94.94%	94.88%	97.80 MB
ConvNeXt-Tiny	Modern ConvNet	28.60	96.69%	96.65%	109.50 MB
Swin-Transformer-Tiny	Hierarchical ViT	28.30	97.23%	97.20%	112.40 MB
Vision-Mamba-Tiny (Vim)	1D Visual SSM	5.50	96.22%	96.18%	22.10 MB
🚀 LiteRS-Mamba (Proposed FP32)	Omnidirectional SSM	2.05	97.19%	97.18%	8.25 MB
⚡ LiteRS-Mamba (Quantized INT8)	Edge Quantized SSM	2.05	97.14%	97.10%	8.58 MB

Efficiency Takeaway: LiteRS-Mamba matches the accuracy of Swin-Transformer ($97.19%$ vs $97.23%$) while requiring $13.8\times$ fewer parameters and $13.1\times$ less storage!

---

💻 Hardware Benchmarking & Pareto Frontier

   High Accuracy
        ^
 98% +  |           ★ LiteRS-Mamba (97.19%, 8.25 MB) [Pareto Optimal]
        |               ● Swin-T (97.23%, 112 MB)
 96% +  |       ● ConvNeXt-T (96.69%, 109 MB)
        |       ● Vision-Mamba (96.22%, 22 MB)
 94% +  |   ● ResNet-50 (94.94%, 98 MB)
        +---------------------------------------------------->
       0 MB           40 MB          80 MB         120 MB
                         Storage Footprint

• CPU Throughput (ONNX Runtime): $297.73\text{ ms}$ ($3.4\text{ FPS}$, a $1.47\times$ acceleration over native PyTorch CPU).
• GPU Throughput (CUDA T4): $155.90\text{ ms}$ ($6.4\text{ FPS}$).
• Quantization Fidelity Audit ($N=350$ scenes): Accuracy delta is merely $0.29%$ ($97.43% \rightarrow 97.14%$) with $99.71%$ decision agreement.

---

🚀 Quickstart: Running the Web App Locally

1. Clone the Repository

git clone https://github.com/mahfuj735/Remote-sensing-Image-Classification.git
cd Remote-sensing-Image-Classification

2. Install Dependencies

pip install -r requirements.txt

3. Launch the Interactive Dashboard

python app.py

Open your browser at http://localhost:7860. You can drag-and-drop satellite scenes, view instant predictions, toggle TTA radiometric alignment, and inspect Grad-CAM attention heatmaps!

---

🌐 Deploying to Hugging Face Spaces (in 2 Minutes)

This repository is structured to run directly on Hugging Face Spaces without any modifications:

1. Create a free account at huggingface.co.
2. Click New Space $\rightarrow$ Set Space Name (e.g. liters-mamba-bd).
3. Select SDK: Gradio and Hardware: CPU Basic (Free).
4. Push this repository to Hugging Face:

   git remote add space https://huggingface.co/spaces/YOUR_USERNAME/liters-mamba-bd
   git push space main
   

5. Your live web application will be active at https://huggingface.co/spaces/YOUR_USERNAME/liters-mamba-bd!

---

📦 Repository Structure

Remote-sensing-Image-Classification/
├── app.py                     # Main interactive Gradio application
├── requirements.txt           # Runtime dependencies
├── README.md                  # Hugging Face + GitHub documentation
├── checkpoints/               # Trained model weights (< 12 MB each)
│   ├── best_liters_mamba_ArcGIS_Map.pth  # Primary champion model (8.25 MB)
│   ├── best_liters_mamba_Bing_Maps.pth   # Bing Maps model
│   ├── best_liters_mamba_Google_Maps.pth # Google Maps model
│   ├── liters_mamba_fp32.onnx            # Serialized ONNX Opset 17
│   └── liters_mamba_int8.onnx            # Post-training INT8 quantized
├── sample_images/             # Test gallery tiles across all 7 LULC classes
│   ├── 01_Agricultural_Land.png
│   ├── 02_Beaches_Coastline.png
│   ├── 03_Built_up_Areas.png
│   ├── 04_Grasslands_Open_Land.png
│   ├── 05_Hills_Mountains.png
│   ├── 06_Trees_Forests.png
│   ├── 07_Water_Bodies.png
│   ├── 08_Google_Agricultural_Land_CrossPlatform.png
│   └── 09_Google_Water_Bodies_CrossPlatform.png
└── assets/                    # Figures, architecture diagrams, and charts
    ├── LiteRS_Mamba_Exact_Architecture.svg
    ├── fig_edge_pareto_efficiency_frontier.png
    ├── fig_edge_latency_fps_comparison.png
    └── fig_tta_visual_gallery_7classes.png

---

📜 Citation & Attribution

If you use this model or code in your research, please cite:

@mastersthesis{sheikh2026litersmamba,
  author       = {Md. Mahfujul Karim Sheikh},
  title        = {Multi-Platform Remote Sensing Land Cover Classification in Bangladesh: Benchmarking Deep Vision Paradigms, Proposing LiteRS-Mamba, and Test-Time Spectral Radiometric Adaptation},
  school       = {Department of Computer Science and Engineering},
  year         = {2026},
  note         = {ID: CSE11220120735}
}

---

📄 License

This project is licensed under the MIT License.