saef-api / MODELS.md
m4az19's picture
Deploy SAHAB-1 backend and pretrained AI
124816b verified
|
Raw History Blame Contribute Delete
2.21 kB
# SAHAB-1 models for m4az19/saef-api
Both ground models download during the Docker build. Do not manually upload
weights, rename the old SAEF `best.pt`, or use BiomedCLIP with this application.
## Satellite water segmentation
- Repository: https://huggingface.co/giswqs/s2-water-unetplusplus-efficientnet-b4
- Revision: `efb6f636480188f48779b41f2a94bb941fd7b900`.
- Architecture: UNet++ / EfficientNet-B4, approximately 19 million parameters.
- Input: Sentinel-2 bands B2, B3, B4, B8, B11 and B12.
- Output: surface-water segmentation, used with acquisition time and valid pixels.
- Downloaded files: `model.pth`, `config.json`, upstream `README.md`.
- License: CC BY 4.0; retain upstream attribution.
- The Dockerfile calls `backend.sources.install_model()` to fetch these files.
- This model does not classify drought severity or predict future flooding.
## Grounded dashboard explanations
- Runtime model: `qwen3.5:0.8b`.
- Model page: https://ollama.com/library/qwen3.5:0.8b
- Runtime: Ollama `0.40.13`, bound to `127.0.0.1:11434` inside the Space.
- License: Apache 2.0 (model).
- Download: `ollama pull qwen3.5:0.8b`, executed by the Docker build supervisor.
- Selects supplied evidence IDs; the server renders actual measured/source values.
- The uploaded manifest records the laptop's previously installed artifact.
The Ollama tag is pulled from upstream at build time and may change; this
deployment does not assert that each cloud build reproduces that same digest.
- Raw Hugging Face safetensors are not a drop-in replacement for this Ollama route.
## ESP32 camera context (not hosted in the Space)
- Component: `espressif/imagenet_cls` version `0.4.2`.
- Model: `espressif-mobilenetv2-s8-v1`, INT8, ESP32-S3.
- Source: https://components.espressif.com/components/espressif/imagenet_cls/versions/0.4.2
- Model size: 3,612,288 bytes, embedded by the pretrained firmware build.
- General ImageNet labels provide camera context; sensor readiness uses engineering
checks. Neither output independently establishes drought impacts.
NDVI and SPI-3 are calculations from environmental inputs, not additional model
weight files. The Space also runs the existing API, acquisition and analysis jobs.