Smart Agriculture with Internet of Things for Precise Crop Prediction using Interfused Machine Learning and Advanced Stacking Ensemble from Soil Parameters

Information about large and remote agricultural regions can be gathered using Internet of Things (IoT) systems, and machine learning approaches can be applied to predict crops. Crop recommendations are determined by factors such as rainfall, moisture, temperature, nitrogen (N), phosphorus (P), potassium (K), pH, and temperature. The dataset contains 2,200 instances and 8 features, leading to suggestions for approximately 22 different crops based on various combinations of these 8 attributes. Using artificial intelligence algorithms in WEKA, the most effective model is developed through supervised learning. The integration of IoT technology has significantly improved crop prediction accuracy by providing real-time soil data. This study further investigates the use of fused machine learning techniques and enhanced stacked ensemble approaches to increase crop prediction accuracy, using soil characteristics gathered from IoT sensors. Due to the complexity and variability of soil conditions, existing crop prediction models often face challenges that result in insufficiently precise forecasts. Existing models may fail to capture temporal relationships and overlook the intricate interactions between soil features. To address these challenges, researchers propose a novel approach that combines multiple machine learning algorithms such as Bidirectional LSTM (BiLSTM), Vanilla Recurrent Neural Networks (VRNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU). To enhance precision, this approach integrates the predictive strengths of these models. The aim of this research is to develop an accurate prediction model that optimizes resource utilization and productivity in agriculture. The stacked ensemble approach achieved a Mean Squared Error (MSE) of 0.045 and an R-squared (R²) value of 0.92, compared to individual models with MSEs ranging from 0.065 to 0.085 and R² values ranging from 0.85 to 0.90. These results demonstrate a significant improvement in prediction accuracy.