SOIL QUALITY FORECASTING WITH OPTIMAL FEATURE SELECTION AND EXTENDED CROSS-STAGE PYRAMID NETWORK

Abstract

Soil quality is essential for sustainable agriculture. Nonetheless, inadequate irrigation methods, improper fertilizer use, and over-cultivation reduce soil quality, thereby decreasing soil fertility. Precise soil quality prediction is crucial for improving agricultural practices. Conventional deep learning models often encounter problems related to superfluous features, high computational demands, and inaccurate predictions. In this work, an innovative deep learning architecture integrating optimal feature selection was proposed to address these challenges. Initially, the Adaptive Parrot Optimization (AdPo) method was used to identify the most relevant features from pre-processed soil data. The Extended Cross Stage Pyramid Network (ExCSP_Net) was introduced to improve soil quality prediction. This network integrates a gated recurrent unit (GRU)-based attention module into the main pathway of the Cross Stage Partial (CSP) model to capture long-range dependencies and emphasize relevant information. In addition, a stacked autoencoder was incorporated before the feature-sharing stage in the short path of the CSP model to reduce dimensionality and generate meaningful representations. The AdPo+ExCSP_Net model demonstrated outstanding performance, achieving an accuracy of 98.58 %, recall of 98.09 %, precision of 98.32 %, F1-score of 98.15 %, Mean Absolute Error (MAE) of 0.53, Root Mean Square Error (RMSE) of 0.65, and coefficient of determination (R²) of 0.99. These findings highlight the effectiveness of the proposed methodology for accurate soil quality prediction and the promotion of sustainable agricultural practices.