FT-TRANSFORMER-DRIVEN CROP SUITABILITY PREDICTION USING MULTI-NUTRIENT INDICES FOR PRECISION AGRICULTURE

This study proposes an intelligent crop decision support framework that integrates soil nutrient profiling with machine learning and deep learning models to enable precision-oriented agricultural planning. Multiple models, including Random Forest, K-Nearest Neighbors, Dense Neural Networks, 1D Convolutional Neural Networks, and the FT-Transformer, are evaluated for their effectiveness in learning non-linear relationships from tabular soil nutrient data. Comparative evaluation indicates that the FT-Transformer achieves the highest predictive accuracy of 98.64%, demonstrating superior capability in capturing multi-nutrient interdependencies. Random Forest attains an accuracy of 92.00%, while KNN records 90.00%, reflecting limitations of distance-based learning in high-dimensional nutrient spaces.

本研究提出了一种融合土壤养分图谱与机器学习、深度学习模型的智能作物决策支持框架，以支撑精准化农业生产规划。针对表格型土壤养分数据中的非线性关系学习效能，本研究对随机森林（Random Forest）、K近邻（K-Nearest Neighbors）、全连接神经网络（Dense Neural Networks）、一维卷积神经网络（1D Convolutional Neural Networks）以及FT-Transformer在内的多种模型展开了评估。对比评估结果表明，FT-Transformer实现了98.64%的最高预测准确率，展现出捕捉多养分相互依存关系的卓越能力。随机森林的准确率达92.00%，K近邻则为90.00%，这反映出基于距离的学习方法在高维养分空间中存在局限性。