Deep Learning for Spatiotemporal Soil Health Prediction in Punjab, Pakistan: A Google Earth Engine-Based CNN-LSTM Framework
Keywords:
Machine Learning/ Deep Learning, CNNs, LSTM, GEE, hybrid CNN-LSTM, Soil Health Prediction, Google Earth Engine, Sentinel-2 Spatio Temporal study, Soil Grids labelsAbstract
Soil degradation in Punjab, Pakistan—a region critical to national food security—threatens the livelihoods of 12 million smallholder farmers due to nutrient mismanagement and rising salinity. This study pioneers the integration of CNN and LSTM architectures for spatiotemporal soil health prediction in Punjab, leveraging multi-sensor data to address regional agro-climatic challenges. This study integrates Sentinel-2 temporal composites, SoilGrids labels, and a hybrid CNN-LSTM model within Google Earth Engine (GEE) to predict soil health indicators (pH, organic matter, NPK). By analyzing 12-month temporal sequences of multispectral data, the model achieved 94% accuracy in classifying soil quality, outperforming traditional methods (Random Forest: 82%, XGBoost: 85%). A GIS-based soil health map highlights critical degradation zones in central Punjab (28% with organic matter <1.5%), enabling 15–25% fertilizer cost reductions through precision agriculture. This study showcases an output of the framework processing 1.2TB of imagery in 5 hours on GEE, demonstrating scalability for arid agro-ecosystems globally, which would never have been possible without a streamlined approach.
