SoilSense: Soil quality prediction using real-time LSTM-GNN

Soil quality is an essential resource for both carbon sequestration and food security. To provide a cloud-native framework for real-time soil quality prediction based on AWS SageMaker, this work combines Long Short-Term Memory (LSTM) networks for temporal modelling with Graph Neural Networks (GNN) for spatial reasoning. Multivariate time-series data, such as soil characteristics (pH, electrical conductivity, organic carbon, nitrogen, phosphorus, potassium, moisture, temperature), and meteorological factors (temperature, rainfall, crop rotation), are collected using automated AWS Lambda to S3 pipelines. Spatial coordinates and temporal embeddings from LSTM encoders are combined to create geographic graphs for GNN processing. A multilayer perceptron is used to calculate the overall Soil Quality Index (SQI). The solution includes automatic cleaning, retraining, drift tracking, and feature engineering. The hybrid LSTM-GNN achieves high accuracy on Kurukshetra data (MSE=0.00023, RMSE=0.015, MAE=0.011, R2=0.956), allowing for data-driven agricultural decision-making and halting soil degradation.