This paper presents a deep learning framework for predicting agricultural land suitability and assessing soil quality, tailored for precision agriculture. We implement six advanced architectures—Fully Connected Neural Networks (FCNN), Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory (Bi-LSTM), Bidirectional Gated Recurrent Units (Bi-GRU), TabNet, and TabTransformer using random splits to capture both spatial and structured dependencies from agricultural data. To address initial data limitations, a 2,000-instance dataset was augmented to 10,000 samples using SMOTE, achieving balanced representation across four classes: High Potential, Moderate Potential, Low Potential, and Requires Attention. Data preprocessing steps included label encoding, feature scaling, and class balancing. Experimental results show TabTransformer achieved the highest accuracy (97.40%), surpassing FCNN (96.60%) and Bi-GRU (96.40%). We additionally report model-averaged LODO results with bootstrap confidence intervals; the large gap vs. random splits exposes spatial dependence and underscores the need for spatially informative features. Despite strong performance, transformer-based models require more computational resources and larger datasets. This study demonstrates the efficacy of deep learning for structured agricultural data and proposes future work in hybrid architectures, adding spatially informative features and IoT-based real-time applications for enhanced decision-making in smart farming.