Learning Soil Texture Fractions Via Pix2Pix Conditional Gans And Geo-Covariates

Abstract

This paper presents a novel application of Pix2Pix conditional Generative Adversarial Networks (cGANs) for pre- dicting soil texture fractions from environmental covariates using ISRO/NRSC Indian Soil Datasets. The problem being investigated is the challenge of large-scale soil mapping for precision agri- culture, where traditional interpolation methods fail to capture complex spatial relationships in sparse government survey data. We employed an enhanced Pix2Pix architecture with composi- tional data constraints, spectral normalization, and spatial cross- validation to ensure proper handling of soil texture fractions that must sum to unity. Our experimental setup utilized 5-kilometer gridded data across the Indian subcontinent with 433,750 pixels, implementing spatial block-based cross-validation to prevent data leakage. The model architecture comprises a 7.08M parameter U-Net generator and 2.77M parameter PatchGAN discriminator, trained with enhanced loss functions including L1, SSIM, and compositional constraints. Our approach achieved 46% improve- ment in L1 loss over 50 epochs, successfully generating soil fraction maps within proper [0,1] bounds while maintaining 80% Apple Silicon GPU utilization. This work demonstrates the at each spatial location, where i=1  I first successful application of conditional GANs to ISRO soil datasets, providing a robust framework for government-scale environmental monitoring and precision agriculture applications.