IRRISIGHT: A Multimodal Remote Sensing Dataset for Irrigation Mapping

IRRISIGHT is a large-scale, multimodal remote sensing dataset for irrigation classification, soil-water mapping, and agricultural monitoring. It spans over 20 U.S. states and provides more than 1.4 million ML-ready georeferenced patches with structured text prompts derived from soil, hydrology, land use, and climate data.

---

🌍 Overview

This repository supports training and evaluation of irrigation mapping models using:

• Sentinel-2 RGB and vegetation indices
• Crop/land/soil/geospatial metadata
• Vision-language prompts
• Supervised and semi-supervised pipelines

---

📦 Hugging Face Dataset

IRRISIGHT dataset is hosted at:
🔗 https://huggingface.co/datasets/OBH30/IRRISIGHT

Each state directory (e.g., Arizona/) contains:

• metadata.jsonl: Structured metadata and text prompts
• *.tar: WebDataset shards with .npy patches and .json attributes

To load:

from datasets import load_dataset
ds = load_dataset("NibirMandal/IRRISIGHT", split="train", streaming=True, data_dir="Georgia")
sample = next(iter(ds))
print(sample["__key__"], sample["meta.json"])

---

📁 Repository Structure

IRRISIGHT/
├── Data/                          # WebDataset .tar files and metadata
│   ├── Arizona_0000.tar
│   └── Arizona/metadata.jsonl
├── data/
│   ├── data_module.py             # batch preparation
│   └── dataset_v2.py              # dataset definition
├── config/
│   ├── supervised_training_gpu.yaml
│   └── dumps/                     # teacher/student configs
├── model_v3/                      # CLIP, RemoteCLIP, BLIP, SAM, KIIM, etc.
├── training_v2/                   # Training, label generation, evaluation
├── utils/, Evaluation.ipynb, validation.ipynb
├── requirements.txt, environment.yml
└── README.md

---

⚙️ Installation

1. Clone the repository

git clone https://github.com/Nibir088/IRRISIGHT.git
cd IRRISIGHT

2. Create and activate environment

conda env create -f environment.yml
conda activate irrisight

3. Install Git LFS (for large files)

git lfs install

---

🚀 How to Run

🏋️ Train a supervised model

python training_v2/Training_Teacher_Model-gpu.py hydra.run.dir=outputs/ experiment=supervised_training_gpu

✨ Generate pseudo-labels for unlabeled data

python training_v2/Label_Generator_Unlabeled.py +experiment=generate_labels

📊 Evaluate model performance

python training_v2/Evaluate_Unlabeled.py +experiment=eval_unlabeled

---

📊 Sample Benchmark

Model	Modalities	Flood	Sprinkler	Drip
ResNet	RGB	35.2	92.2	88.5
SegFormer	RGB	86.2	91.7	85.9
CLIP	RGB + Text Prompt	90.1	93.1	90.7
RemoteCLIP	RGB + Text Prompt	90.9	93.7	92.3
KIIM	RGB + Crop + Land	93.6	95.8	94.6

---

📸 Dataset Visuals

Figure	Description	Image
1	🔍 Sample data patch with multimodal inputs including RGB, vegetation indices, and masks.
2	🛠️ End-to-end data processing pipeline for converting raw geospatial data to ML-ready patches.
3	📊 Distribution of model confidence scores across different U.S. states during weak supervision.
4	🧪 Evaluation framework used to test model generalizability across held-out states.

---

📝 License

Released for academic research use only. Contact authors for commercial use.

---

📬 Contact

Maintainer: @Nibir088
Dataset: OBH30/IRRISIGHT on Hugging Face