Papahaven/DSSA-LUCAS

Dual Surface-Subsurface Adapter (DSSA)

Official pre-trained model weights for Dual Surface-Subsurface Adapter (DSSA) for multimodal land-cover classification using ground-level RGB photography and topsoil physicochemical properties.

• GitHub Codebase: https://github.com/siam4201/DSSA-LUCAS
• Hugging Face Model Hub: https://huggingface.co/Papahaven/DSSA-LUCAS

---

1. Models Included

Model Variant	Parameters	Latency (B=1)	Throughput (B=64)	Peak VRAM	Test Accuracy	Macro F1	Cohen's Kappa
DSSA-Lite (Edge)	3.99M	9.85 ms	3,724.4 img/s	698.0 MB	80.63%	0.7487	0.7553
DSSA-Standard (Proposed)	8.42M	13.87 ms	3,344.9 img/s	893.7 MB	86.20%	0.8287	0.8286

• best_dssa_full_standard.pth (32.5 MB):

  • Full proposed architecture (8.42M parameters).
  • Integrates Spatial Attention Decomposition, Physics-Guided Modality Relevance (PGMR) routing, and Adaptive Zero-Soil canopy gating.
  • Achieves 86.20% Accuracy and 0.8287 Macro F1 on the frozen LUCAS benchmark (N = 2,917).

• dssa_lite_best_seed42.pth (15.5 MB):

  • Edge-optimized architecture (3.99M parameters, 52.6% parameter reduction).
  • Truncates EfficientNet-B0 at Stage 5 and discards the 1,280-channel expansion head, keeping adapter overhead down to just 528k parameters.
  • Optimized for real-time edge surveying (<10 ms latency, 698 MB VRAM).

---

2. Hardware Environment & Training Compute

All models were trained and profiled under the following hardware environment:

• GPU: NVIDIA GeForce RTX 5070 Ti (16 GB GDDR7 VRAM, Blackwell SMs)
• CPU: x86-64 Host Processor
• Software: PyTorch 2.6.0, CUDA 12.8, cuDNN 9.x, FP32 Single Precision

Training Durations (Approximate)

• DSSA-Standard:
  • ~45–55 seconds per epoch (batch size = 64, 13,608 training samples).
  • Total Training Time: ~20–25 minutes (converged in 25–28 epochs with early stopping patience = 15).
• DSSA-Lite:
  • ~25–35 seconds per epoch.
  • Total Training Time: ~10–12 minutes (converged in 20–24 epochs).

---

3. Quick Inference Examples

A. Run Inference with Proposed DSSA-Standard (8.42M)

import torch
from huggingface_hub import hf_hub_download
from models.dssa_model import DSSAModel

# 1. Download checkpoint
ckpt_path = hf_hub_download(repo_id="Papahaven/DSSA-LUCAS", filename="best_dssa_full_standard.pth")

# 2. Instantiate DSSA-Standard
feature_cols = ['pH_H2O', 'pH_CaCl2', 'OC', 'CaCO3', 'N', 'P', 'K', 'EC']
visible_cols = ['pH_H2O', 'pH_CaCl2', 'OC', 'CaCO3', 'EC']
subsurface_cols = ['N', 'P', 'K', 'EC']

model = DSSAModel(
    feature_cols=feature_cols,
    visible_cols=visible_cols,
    subsurface_cols=subsurface_cols,
    num_classes=6,
    shared_dim=256,
    num_heads=4,
    use_spatial_attention=True,
    use_physics_guidance=True,
    use_soil_gating=True,
    pretrained=False,
)

state_dict = torch.load(ckpt_path, map_location="cpu", weights_only=True)
model.load_state_dict(state_dict)
model.eval()

# 3. Predict: (B, 3, 224, 224) RGB image and (B, 8) normalized soil vector
img = torch.randn(1, 3, 224, 224)
soil = torch.randn(1, 8)
with torch.no_grad():
    logits = model(img, soil)
pred = logits.argmax(dim=-1).item()
print("DSSA-Standard Predicted Class:", pred)

B. Run Inference with Edge-Optimized DSSA-Lite (3.99M)

import torch
from huggingface_hub import hf_hub_download
from models.dssa_lite import DSSALiteModel

# 1. Download checkpoint
ckpt_path = hf_hub_download(repo_id="Papahaven/DSSA-LUCAS", filename="dssa_lite_best_seed42.pth")

# 2. Instantiate DSSA-Lite
model_lite = DSSALiteModel(
    num_classes=6,
    d=128,
    num_heads=4,
    pretrained=False,
    dropout=0.15,
)

state_dict = torch.load(ckpt_path, map_location="cpu", weights_only=True)
model_lite.load_state_dict(state_dict)
model_lite.eval()

# 3. Predict: (B, 3, 224, 224) RGB image and (B, 8) normalized soil vector
img = torch.randn(1, 3, 224, 224)
soil = torch.randn(1, 8)
with torch.no_grad():
    logits = model_lite(img, soil)
pred = logits.argmax(dim=-1).item()
print("DSSA-Lite Predicted Class:", pred)

---

4. Citation

@article{dssa2026,
  title={Dual Surface-Subsurface Adapter for Ground-Level Land Cover Classification with Soil Physicochemical Context},
  author={Siam, Sarker and Mim, Farhana Nur},
  journal={IEEE Transactions / Conference Submission},
  year={2026},
  url={https://github.com/siam4201/DSSA-LUCAS}
}