RTCNet: A Robust Hybrid Deep Learning Model for Soil Property Prediction Under Noisy Conditions

Accurate prediction of soil fertility and soil organic carbon (SOC) plays a critical role in precision agriculture and sustainable soil management.However, the high spatial-temporal variability inherent in soil properties, compounded by the prevalence of noisy data in real-world conditions, continues to pose significant modeling challenges.To address these issues, a robust hybrid deep learning model, termed RTCNet, was developed by integrating Recurrent Neural Networks (RNNs), Transformer architectures, and Convolutional Neural Networks (CNNs) into a unified predictive framework.Within RTCNet, a one-dimensional convolutional layer was employed for initial feature extraction, followed by MaxPooling for dimensionality reduction, while sequential dependencies were captured using RNN layers.A multi-head attention mechanism was embedded to enhance the representation of inter-variable relationships, thereby improving the model's ability to handle complex soil data patterns.RTCNet was benchmarked against two conventional models-Artificial Neural Network (ANN) optimized with a Genetic Algorithm (GA), and a Transformer-CNN hybrid model.Under noise-free conditions, RTCNet achieved the lowest Mean Squared Error (MSE) of 0.1032 and Mean Absolute Error (MAE) of 0.1852.Notably, under increasing noise levels, RTCNet consistently maintained stable performance, whereas the comparative models exhibited significant performance degradation.These findings underscore RTCNet's superior resilience and adaptability, affirming its utility in field-scale agricultural applications where sensor noise, data sparsity, and environmental fluctuations are prevalent.The demonstrated robustness and predictive accuracy of RTCNet position it as a valuable tool for optimizing nutrient management strategies, enhancing SOC monitoring, and supporting informed decision-making in sustainable farming systems.