Update model card for EuroSAT Field Scout

README.md

@@ -1,23 +1,45 @@
 ---
 language: en
 license: mit
+library_name: pytorch
 tags:
   - pytorch
   - image-classification
   - satellite-imagery
   - computer-vision
   - eurosat
+  - small-models
+  - local-first
 datasets:
-  - eurosat
+  - torchgeo/eurosat
 metrics:
   - accuracy
 pipeline_tag: image-classification
 ---
 
-# SimpleNet — EuroSAT Land-Use Classifier
+# SimpleNet EuroSAT Land-Use Classifier
+
+Small PyTorch CNN for 10-class EuroSAT land-use classification.
+
+This model is used by the Gradio Space:
+
+- Hackathon Space: https://huggingface.co/spaces/build-small-hackathon/EuroSATFieldScout
+- Personal Space: https://huggingface.co/spaces/yava-code/eurosat-field-scout
+
+## Model Details
+
+| Field | Value |
+| --- | --- |
+| Architecture | Four Conv-BN-ReLU-Pool blocks plus dense classifier |
+| Parameters | 2,492,170 |
+| Input | RGB image resized to 64 x 64 |
+| Output | 10 EuroSAT land-use classes |
+| Runtime | CPU-friendly PyTorch inference |
+
+Classes:
 
-Lightweight CNN (~850K params) trained from scratch on the EuroSAT dataset
-for 10-class satellite image classification.
+`AnnualCrop`, `Forest`, `HerbaceousVegetation`, `Highway`, `Industrial`,
+`Pasture`, `PermanentCrop`, `Residential`, `River`, `SeaLake`
 
 ## Usage
 
@@ -25,25 +47,34 @@ for 10-class satellite image classification.
 import torch
 from huggingface_hub import hf_hub_download
 
-# Download
-weights = hf_hub_download(repo_id="yava-code/eurosat-simplenet", filename="best_model.pth")
-
-# Load (you need model.py from this repo)
 from model import SimpleNet, CLASS_NAMES
-model = SimpleNet()
+
+weights = hf_hub_download(
+    repo_id="yava-code/eurosat-simplenet",
+    filename="simple_net_v1.pth",
+)
+
+model = SimpleNet(num_classes=len(CLASS_NAMES))
 model.load_state_dict(torch.load(weights, map_location="cpu"))
 model.eval()
 ```
 
-## Architecture
-4 convolutional blocks (Conv→BN→ReLU→Pool) + FC classifier.
-Channels: 3→32→64→128→256. Spatial: 64→32→16→8→4.
+## Training Notes
+
+- Dataset: EuroSAT RGB land-use imagery
+- Optimizer: Adam
+- Augmentations: flips, rotations, color jitter
+- Goal: compact educational classifier, not production remote-sensing model
+
+## Deployment Notes
+
+For the Build Small Hackathon Space, the same state dict is converted to
+float16 chunks so the app can be reviewed without Git LFS write permissions in
+the organization repo. The runtime reconstructs the bytes, casts tensors back to
+float32, and loads the explicit `SimpleNet` architecture.
 
-## Training
-- **Dataset**: EuroSAT (27,000 images, 10 classes)
-- **Optimizer**: Adam (lr=1e-3, StepLR γ=0.1 every 5 epochs)
-- **Augmentations**: Flip, Rotation ±15°, ColorJitter
-- **Epochs**: 20
+## Limits
 
-## Demo
-👉 [Try the live demo on Spaces](https://huggingface.co/spaces/yava-code/eurostat)
+This model is trained for EuroSAT-style RGB tiles. It should not be used for
+production geospatial decisions without target-domain validation, calibration,
+and uncertainty handling.