Predicting soil moisture under dynamic climate conditions is challenging due to intricate dependencies within the data. This study presents a surrogate deep learning (SDL) model with a multitask learning (MTL) approach to improve daily soil moisture predictions across spatiotemporal scales. The model employs a two-level encoding process, first compressing climate parameters into a single feature and then applying sequential encoding to capture long-term temporal patterns within a one-year timeframe for better generalization. Seasonality detection using autocorrelation facilitates data resampling into homogeneous samples, enhancing the SDL model by optimizing hyperparameters through efficient weight sharing between layers. To evaluate the effectiveness of MTL, three different SDL architectures such as LSTM, ConvLSTM, and BiLSTM were implemented for a comprehensive analysis. All models struggle with soil moisture prediction, particularly during dry periods, where LSTM experiences the most significant accuracy drop. While BiLSTM demonstrates better performance, its effectiveness remains constrained. However, integrating MTL enhances model stability and spatiotemporal accuracy, reducing errors across various conditions and achieving a 10% improvement due to better data representation, enabling SDL models to capture regional heterogeneity more effectively.