DeepAR



---

⚠️ Note: This is an ongoing project under active development. Features, documentation and code may change.

---

DeepAR is a deep learning model designed for Atmospheric Rivers (AR) detection and segmentation from climate data (using the Climate variable IVT, IVT_u, IVT_v). It utilizes a modified, prompt-less Segment Anything Model (SAM) to generate AR masks.

Model Architecture

The DeepAR model processes data through a three-stage pipeline:

1. Input Generator (IVT2RGB): A CNN that converts 3 channel climate data (Integrated Vapor Transport: ivt, ivtu, ivtv) into a 3 channel RGB-like image suitable for the image encoder.
2. Segmentation (SamAR): A modified SAM model that operates without prompts. It uses a learned no_mask_embedding to generate segmentation masks from the features produced by the image encoder.

The diagram below illustrates the architecture: 

Below is the architecture of the `IVT2RGB` module: 

Data

The model is designed to work with NetCDF files (.nc) containing ivt, ivtu and ivtv variables. Use the dataset class ARInferenceDataset for loading and preprocessing the data during inference.

Installation

1. Clone the repository:

   git clone https://github.com/mphysicus/deep_AR.git
   cd deep_AR

2. Install the package:

   pip install -e .

Pretrained Models

✨Coming Soon✨We will be uploading pre-trained model weights soon.

🙏 Acknowledgment

We are thankful to Segment Anything for releasing their code as open-source contributions. We are also thankful to AdaLoRA for their open-source contribution in AdaLoRA PEFT training.