Soil degradation due to climate change, nutrient depletion, and poor management endangers long-term agricultural productivity. Traditional soil testing is reliable; however, it is time-consuming and does not provide real-time feedback for precision farming. This study presents an integrated IoT-Machine Learning system for the early detection and prediction of soil health, utilizing a 7-in-1 multi-sensor array. The system monitors seven important parameters (pH, nitrogen, phosphorus, potassium, electrical conductivity, temperature, and humidity), assessed using supervised Machine Learning (ML) models (Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and Neural Network (NN)). Training and evaluation were conducted using a hybrid dataset that combined Kaggle's Crop Recommendation data with field measurements. DT and RF achieved high accuracies (97.57% and 94.3%), with SHAP analysis confirming phosphorus, nitrogen, and potassium as the dominant determinants. NN established nonlinear connections (93.3%) but lacked interpretability, whereas SVM performed poorly (70.83%). The framework illustrates how explainable IoT-ML systems can provide actionable, transparent soil-health information, enabling data-driven precision agriculture decisions.