SatelliteImage_CloudSegmentation_Unet

🛰️ Overview

The SatelliteImage_CloudSegmentation_Unet is a U-Net based model designed for semantic segmentation of satellite imagery. Its purpose is to accurately classify every pixel in an input image as either "Cloud" or "Background/Clear Sky." This is critical for pre-processing Earth Observation (EO) data before tasks like land cover mapping or atmospheric correction.

🧠 Model Architecture

The model employs the classic U-Net architecture, which is highly effective for biomedical and remote sensing segmentation due to its symmetric encoder-decoder structure with skip connections.

• Encoder (Contracting Path): Consists of repeated convolutional and pooling layers to capture contextual information and build high-level feature maps.
• Decoder (Expanding Path): Uses up-convolutional layers to increase the resolution of the feature maps.
• Skip Connections: Directly connect feature maps from the encoder to the corresponding layers in the decoder. This is vital for preserving fine-grained details needed for precise boundary localization.
• Input: RGB satellite image patches of size 256x256.
• Output: A 256x256 pixel-wise mask with 2 channels, representing the probability distribution for the two classes (Cloud and Background).

🎯 Intended Use

This model is intended for use in remote sensing and geospatial applications:

1. EO Data Pre-processing: Automatically generating masks to filter out cloudy regions, ensuring the reliability of subsequent land-use classification or agricultural monitoring.
2. Atmospheric Science: Quantifying cloud fraction and distribution over large geographic areas for climate modeling.
3. Disaster Response: Quickly assessing the visibility of ground features (e.g., flood extent) after a weather event.

⚠️ Limitations

1. Thin/Cirrus Clouds: The model may struggle with very thin, semi-transparent cirrus clouds, often misclassifying them as clear sky due to low contrast.
2. Shadows: Cloud shadows on the ground can sometimes be mistakenly classified as cloud due to their low brightness values.
3. Resolution Dependence: Trained on 256x256 patches. Applying the model directly to very high-resolution images (e.g., 4k) without appropriate tiling and handling may lead to boundary artifacts.

---

MODEL 2: Toxicology_StructureToxicity_GNN

This model is a Graph Neural Network (GNN) for predicting chemical toxicity based on molecular graph structure.

config.json

{
  "_name_or_path": "custom-graph-tox-predictor",
  "architectures": [
    "GraphConvolutionalNetwork"
  ],
  "model_type": "molecular_property_prediction",
  "graph_type": "molecular_graph",
  "node_features": 74,
  "edge_features": 12,
  "num_gcn_layers": 3,
  "hidden_dim": 128,
  "global_pooling": "readout_mean",
  "output_dim": 1,
  "task_type": "binary_classification",
  "id2label": {
    "0": "Non-Toxic",
    "1": "Toxic"
  },
  "label2id": {
    "Non-Toxic": 0,
    "Toxic": 1
  },
  "pytorch_version": "2.1.0"
}