🟢 Wetlands Detection from Terrain — ASPP U-Net Workflow (SWEEP Demo)

This SWEEP demo shows how to detect inland wetlands using only terrain-derived variables and a lightweight ASPP U-Net model. It uses freely available DEMs (e.g., USGS 3DEP), aligns binary wetland labels, derives slope/curvature-based inputs, and outputs high-resolution wetland probability maps. All steps are orchestrated by a multi-cloud-ready SWEEP DAG.

🌿 Wetlands are among the most valuable carbon sinks on Earth. Yet they remain hard to map. This workflow helps automate detection across catchments using only elevation data — no satellite imagery required.

---

⚙️ What This Workflow Does

Task	Description
fn_read_hucs	Reads a list of HUC12 watershed IDs from an input source (e.g., GeoJSON or CSV) and passes them downstream for processing.
fn_fetch_stacks	Fetches and stacks terrain-derived rasters (e.g., slope, aspect, TPI, elevation) for each HUC12 and uploads the .tif stack to S3.
predict_huc	Downloads the stacked raster from S3, loads a pretrained U-Net model from Hugging Face, runs patch-based prediction, and uploads the binary wetland mask to S3.
mosaicpredictions	Mosaics all per-HUC binary wetland predictions into a single output raster (e.g., for a watershed, state, or region). Saves to S3.

---

🔁 SWEEP DAG (YAML)

name: wetland_detection_unet
description: Derive terrain & predict wetlands using deep learning
steps:
  - name: fn_read_hucs
    function: fn_read_hucs

  - name: fn_fetch_stacks
    function: fn_fetch_stacks
    depends_on: [fn_read_hucs]

  - name: predict_bucket
    function: predict_bucket
    depends_on: [fn_fetch_stacks]

  - name: mosaicpredictions
    function: mosaicpredictions
    depends_on: [predict_bucket]

---

🧪 Model Details

• Architecture: ASPP U-Net (4-channel input → 1-channel output)
• Inputs: slope, aspect, curvature, elevation
• Loss: Binary Cross Entropy + Dice
• Training: 128×128 patches
• Libraries: torch, rasterio, albumentations, numpy

---

🖼️ Example Output

---

🚀 Quickstart

1. Clone the repo:

git clone https://github.com/SWEEP-Inc/sweep-wetlands.git
cd demo-workflows/wetlands-demo

2. Install dependencies:

pip install -r requirements.txt

3. Deploy with SWEEP:

4. Trigger full DAG with SWEEP CLI or REST API:

sweep run workflow.yaml --inputs input_dem=data/example_dem.tif

---

📈 Results (PNW}

Metric	Mean ± SD
mIoU	0.73 ± 0.06
F1 Score	0.84 ± 0.05
Precision	0.88 ± 0.04
Recall	0.81 ± 0.07

Note: Performance will vary by watershed depending on label quality and topographic contrast.

---

📦 License

MIT License — free to use and modify with credit to DotMote Labs / SWEEP.

---

📬 Contact

DotMote Labs
🌍 https://dotmotelabs.com/sweep.html
📧 hello@sweep.run

Want to use this workflow on your landscape? Need a wetland model developed your region? Get in touch.