rs-dl-architect
rs-dl-architect — Expert guidance for designing, analyzing, and reviewing deep learning architectures in remote sensing and satellite imager
Open source Repository Open in the app JSON README (API)
About
rs-dl-architect — Expert guidance for designing, analyzing, and reviewing deep learning architectures in remote sensing and satellite imagery. Covers segmentation, change detection, multi-modal fusion, temporal analysis, foundation models (SatMAE, Prithvi, GeoSAM), object detection, and uncertainty quantification. Supports both PyTorch and TensorFlow with production-ready code templates.
Details
- Kind
- Plugins
- Topic
- No topic detected
- Publisher
- rupesh4604
- Origin
- marketplace
- Category
- ferramentas
- Stars
- 1
- Last push
- 2026-06-15T12:43:16Z
- Repository state
- ativo
- License
- MIT
- Added
- 2026-08-30 01:48:58
- Updated
- 2026-08-30 01:48:58
- Origin id
rupesh4604/rs-dl-architect/rs-dl-architect
README
# rs-dl-architect **Expert guidance for designing, analyzing, and reviewing deep learning architectures in remote sensing and satellite imagery.**  [](LICENSE) [](https://github.com/anthropics/skills) [](https://your-username.github.io/rs-dl-architect) A comprehensive Claude AI skill covering segmentation, change detection, multi-modal fusion, temporal analysis, foundation models (SatMAE, Prithvi, GeoSAM), object detection, and uncertainty quantification. Supports both **PyTorch** and **TensorFlow** with production-ready code templates. ([ Link to website ](https://rupesh4604.github.io/rs-dl-architect/)) --- ## 🎯 What This Skill Does - **Designs** architectures tailored to your remote sensing task (task-shape diagnosis first, not reflexive model recommendations) - **Analyzes** existing models, papers, and checkpoints (explains what's actually happening, not just listing layers) - **Reviews** your code/architecture with a 10-section rigorous checklist (blocking issues + concrete fixes, not vague suggestions) **Key differentiator:** Not generic computer vision adapted to RS — built from the ground up for satellite/aerial imagery physics, GSD considerations, geographic generalization, and multi-spectral/multi-temporal/multi-modal challenges. --- ## 🚀 Quick Start ### Installation **Option 1: Via Claude Code Plugin** _(Recommended)_ ```bash claude plugin install rs-dl-architect@Rupesh4604 ``` **Option 2: Manual Installation** ```bash # Clone the repository git clone https://github.com/Rupesh4604/rs-dl-architect.git # Copy to Claude skills directory cp -r rs-dl-architect/rs-dl-architect ~/.claude/skills/ # For project-specific installation cp -r rs-dl-architect/rs-dl-architect ./.claude/skills/ ``` **Option 3: Direct Download** Download the [latest .skill file](https://github.com/Rupesh4604/rs-dl-architect/releases) and upload via Claude.ai or Claude Code. ### Usage Just ask Claude naturally — the skill triggers automatically on relevant queries: ``` "I need to detect crop field boundaries from 10m Sentinel-2 with 500 labels" "How should I fuse SAR and optical for flood mapping?" "Review my U-Net — using ImageNet weights with 13 Sentinel-2 bands" "Should I use SatMAE or DINOv2 for EuroSAT land cover classification?" ``` --- ## 📖 Coverage ### Architectures & Tasks | Domain | Coverage | | ------------------------- | ------------------------------------------------------------------------------------------------ | | **Segmentation** | U-Net family, SegFormer, UNetFormer, Mask2Former, DC-Swin, RS-Mamba, boundary-aware patterns | | **Change Detection** | Siamese (FC-Siam-diff), BIT, ChangeFormer, ChangeMamba, semantic CD, continuous monitoring | | **Data Fusion** | SAR+optical, RGB+DEM, hyperspectral+LiDAR, early/mid/late fusion, modality dropout, CMX, DOFA | | **Temporal Analysis** | ConvLSTM, U-TAE, TSViT, Presto, TempCNN, irregular time series, cloud masking | | **Foundation Models** | SatMAE, Prithvi-EO, Clay, DOFA, GeoSAM, RemoteCLIP, LoRA/PEFT adaptation, multispectral adapters | | **Object Detection** | RetinaNet, FCOS, DETR/DINO, oriented bbox (DOTA), small-object patterns, YOLOv8-OBB | | **Interpretability & UQ** | Grad-CAM, MC dropout, deep ensembles, evidential learning, conformal prediction, calibration | ### Code Templates (PyTorch + TensorFlow) - Multispectral input adapters for ImageNet weights - Minimal U-Net with shape annotations - Siamese change-detection (FC-Siam-diff) - Combined Dice + CE loss - Training loop with mixed precision + gradient accumulation - Tile-based inference with cosine-weighted blending - Deep ensemble patterns --- ## 💡 Example Use Cases <details> <summary><strong>Architecture Selection</strong></summary> **Query:** ``` I need to detect crop field boundaries from 10m Sentinel-2 data with only 500 labeled samples. What architecture should I use? ``` **Response:** - Recommends foundation-model fine-tuning (Prithvi-EO with LoRA) - Explains why label efficiency matters - Provides PyTorch implementation with proper multispectral input handling - Lists failure modes (GSD mismatch, small fields <2 pixels) - Suggests alternatives (SatMAE++, from-scratch U-Net as baseline) </details> <details> <summary><strong>Multi-Modal Fusion Design</strong></summary> **Query:** ``` How do I combine SAR and optical data for flood mapping? Should I fuse early or late? ``` **Response:** - Analyzes modality differences (speckle vs additive noise, cloud penetration) - Recommends mid-level cross-attention fusion (not early concat) - Flags modality dropout requirement for robustness - Provides dual-encoder architecture with shape annotations - Cites CMX, DeCUR as reference implementations </details> <details> <summary><strong>Code Review</strong></summary> **Query:** ``` Review my U-Net for building extraction — using ImageNet weights with 13 Sentinel-2 bands via channel concat. ``` **Response:** - **Blocking issue:** Silently dropping spectral information (red-edge, SWIR) - **Fix:** Use multispectral adapter pattern from code templates - Checks receptive field vs GSD (10m → need larger context) - Suggests boundary-aware loss for cadastral precision - Verifies train/val split is region-based, not random tiles </details> <details> <summary><strong>Foundation Model Selection</strong></summary> **Query:** ``` Should I use SatMAE or DINOv2 for land cover classification on EuroSAT? ``` **Response:** - Compares both: SatMAE native multispectral vs DINOv2 stronger features - Recommends SatMAE++ with linear probe baseline first - Provides LoRA fine-tuning template for parameter efficiency - Explains when each wins (SatMAE: <10k labels; DINOv2: RGB-only high-res) - Notes EuroSAT is small (27k tiles) — foundation models mandatory </details> --- ## 🧠 How It Works ### Three-Mode Workflow The skill automatically identifies your intent: 1. **Build Mode** — You want a new architecture or implementation - Task-shape diagnosis (what's being predicted, input modality, GSD, label budget) - Routes to relevant subdomain reference (segmentation, fusion, temporal, etc.) - Produces 5-part recommendation: What / Why / Cost / Failure Modes / Alternatives 2. **Analyze Mode** — You want to understand an existing architecture/paper - Structured walkthrough: data flow → tensor shapes → key decisions → novelty - Comparisons to baselines and related work - Diagrams when branches/attention is non-trivial 3. **Review Mode** — You want critique or debugging help - 10-section checklist (task-architecture match, data flow, loss, evaluation, etc.) - Blocking issues with concrete fixes (not "consider regularization") - Acknowledges strengths + flags weaknesses with evidence ### Cross-Cutting Principles Built into every recommendation: - **Channels ≠ RGB** — multispectral input needs proper adapters, never silent band dropping - **GSD matters** — receptive field, augmentation, class definitions all depend on ground sampling distance - **Geographic generalization** — held-out region tests, not random tile splits - **Foundation models changed the calculus** — for <50k labels, fine-tune beats from-scratch - **Temporal context is often free signal** — multi-date stacks almost always beat single-date - **Class imbalance is the norm** — weighted loss / focal loss / sampling strategies required --- ## 📂 Repository Structure ``` rs-dl-architect/ ├── README.md # This file ├── docs/ # GitHub Pages documentation site │ └── index.html │ └── privacy.html ├── rs-dl-architect/ # The actual skill │ ├── SKILL.md # Main entry point (~120 lines) │ └── references/ # Loaded on-demand │ ├── segmentation.md │ ├── change-detection.md │ ├── data-fusion.md │ ├── temporal-analysis.md │ ├── foundation-models.md │ ├── object-detection.md │ ├── interpretability-uncertainty.md │ ├── review-checklist.md │ ├── code-templates.md │ └── design-principles.md ├── LICENSE └── .gitignore ``` **Size:** ~1,460 lines of expert guidance across 10 files **Main file:** 120 lines (routes to references) **References:** 100–300 lines each (loaded only when relevant) --- ## 🎓 Target Users - Remote sensing researchers - GIS engineers and geospatial AI practitioners - Graduate students in earth observation / geospatial ML - Satellite imagery ML engineers - Anyone working with Sentinel-1/2, Landsat, PlanetScope, aerial imagery, hyperspectral, SAR, or multi-modal EO data **Assumes:** MS/PhD-level technical depth, familiarity with PyTorch or TensorFlow, working knowledge of remote sensing fundamentals. --- ## 🤝 Contributing Contributions welcome! Areas of particular interest: - Additional code templates (e.g., TensorFlow equivalents, JAX patterns) - Newer foundation models (as they're released) - Domain-specific extensions (agriculture, disaster response, urban planning) - More review checklist items - Dataset recommendations and benchmark updates **To contribute:** 1. Fork the repository 2. Create a feature branch (`git checkout -b feature/new-reference`) 3. Make your changes (follow the existing reference structure) 4. Test with realistic queries 5. Submit a Pull Request See [CONTRIBUTING.md](CONTRIBUTING.md) for detailed guidelines. --- ## 📄 License This skill is released under the [MIT License](LICENSE). You are free to: - Use commercially - Modify and redistribute - Use privately - No warranty or liability --- ## 🔗 Links - **Documentation:** [https://Rupesh4604.github.io/rs-dl-architect](https://Rupesh4604.github.io/rs-dl-architect) - **Issues & Discussions:** [GitHub Issues](https://github.com/your-username/rs-dl-architect/issues) - **Privacy Policy:** [rs-dl-architect/privacy](https://rupesh4604.github.io/rs-dl-architect/privacy.html) - **Claude Skills Specification:** [Anthropic Skills Repo](https://github.com/anthropics/skills) - **Related Work:** [Awesome Remote Sensing Change Detection](https://github.com/wenhwu/awesome-remote-sensing-change-detection) --- ## 📊 Stats    **Coverage:** Segmentation • Change Detection • Data Fusion • Temporal • Foundation Models • Detection • Interpretability • Review --- ## 🙏 Acknowledgments Built for the remote sensing and geospatial AI community. Special thanks to: - The teams behind SatMAE, Prithvi-EO, Clay, GeoSAM, and other RS foundation models - Authors of BIT, ChangeFormer, U-TAE, and other open-source RS architectures - The broader earth observation ML research community --- ## 📮 Contact **Maintainer:** [M Rupesh Kumar Yadav](<[@Rupesh4604](https://github.com/Rupesh4604)>) **Email:** rupesh32003@gmai.com **Institution:** IIT Bombay, M.Tech For questions, feedback, or collaboration opportunities, open an issue or reach out directly. --- <p align="center"> <strong>Built with ❤️ for the remote sensing community</strong> <br> <sub>Making deep learning architecture decisions clearer, one satellite image at a time</sub> </p>