Attention Based Conditional GAN for Synthetic Crop Data: Solving Agriculture’s Data Availability and Quality Challenges

Overview

This project aims to generate synthetic data that closely resembles real-world data using Generative Adversarial Networks (GANs). In this implementation, we utilize a custom self-attention layer and introduce a correlation constraint to ensure that the generated data preserves the correlation structure of the real data.

Dataset

In this work we utilized crop production recomendation data which is publically available in Kaggle[1].

Implementation Details

Synthetic Data Generation

GAN Architecture:

• 
• Our framework introduces a multi-head attention layer within the generator. This innovation allows for a more nuanced discernment and emphasis on critical inter-feature relationships in tabular data, significantly improving the model's capability to capture and replicate complex data dynamics in the agricultural domain.

Evaluation

• Though the labels 'pH' and 'Temperature' show almost 95% probability distribution, the number of each label is dramatically various. Moreover, it is the limitation of tabular GANs that they are not able to generate a specific amount of data. Nevertheless, our proposed approach performs better in producing evenly distributed data that closely resembles the original data in all columns with a likeness of 98%, except for the 'P' (Phosphorus) column, where the synthetic data shows a more rapid increase initially, suggesting a higher density of lower values compared to real data .
• 

Model	CC compared to original data
TabGAN[2]	0.631
CTGAN [3]	0.752
Proposed model	0.989

Comparative analysis of Correlation Coefficients (CC) relative to original data for proposed Model and existing State-of-the-Art tabular data generation Methods..

License

This work is licensed under the Custom Collaborative License. Any modification or redistribution of this work requires explicit permission from all collaborators listed in the acknowledgments section of this README file. For inquiries about modifications or collaborations, please contact the project maintainers.Also you can find

Acknowledgments We would like to acknowledge the contributions of the following individuals from the Data Centric AI class:

-M.J. Aashik Rasool

-Abrar Alabdulwahab

-Sevara Mardieva

-Dong Seok Kim

Additionally, we cite the following work that has informed and inspired aspects of this project: Rasool, M. A., Alabdulwahab, A., Mardieva, S., & Kim, D. S. "ATTENTION BASED CONDITIONAL GAN FOR SYNTHETIC CROP DATA: SOLVING AGRICULTURE’S DATA AVAILABILITY AND QUALITY CHALLENGES." They implemented the approach described in this project.

• Link to paper

References

[1] https://www.kaggle.com/datasets/aksahaha/crop-recommendation.

[2] Park, Noseong, et al. "Data synthesis based on generative adversarial networks." arXiv preprint arXiv:1806.03384 (2018).

[3] Xu, Lei, et al. "Modeling tabular data using conditional gan." Advances in neural information processing systems 32 (2019).