CropGAN: A conditional GAN framework for synthetic tabular data augmentation in crop recommendation systems

Abstract

Crop recommendation systems play a crucial role in precision agriculture by enabling informed decisions about which crops to cultivate in a specific location. However, their performance is often hindered by the inherent scarcity, imbalance, and regional bias of tabular agricultural datasets. These limitations reduce the reliability and generalizability of crop recommendation models, especially in data-scarce regions. Existing synthetic data generation methods, such as the Synthetic Minority Oversampling Technique (SMOTE) and variational autoencoders (VAEs), struggle to handle high-dimensional, structured agricultural data with categorical variables. Moreover, existing generative adversarial networks (GANs) are primarily image-focused. This study proposes a re-engineered GAN, termed the Crop Recommendation GAN (CropGAN), to generate tabular, multi-crop recommendation data using a class-conditioning mechanism. The framework was designed to handle complex, non-linear datasets with both numerical and categorical variables, addressing dataset size, imbalance, and limited diversity in multi-crop datasets. CropGAN was trained on a dataset of 5,000 samples with 10 crop classes and evaluated against SMOTE and VAE using statistical data-quality and classification-performance metrics. In terms of data-quality assessment, SMOTE best preserved the original data distributions, while CropGAN introduced greater diversity in the synthetic datasets. In terms of classification performance, models trained on CropGAN-generated samples consistently achieved the highest performance, with the support vector machine (SVM) yielding the best accuracy of 99.4%. This result suggests that adversarial learning can be adapted for tabular agricultural datasets to improve the performance of crop recommendation systems in data-scarce regions.