Soil Quality Prediction using Machine Learning

Abstract

Soil quality prediction (SQP) plays a crucial role in agriculture, environmen- tal management, and civil engineering. Traditional assessment methods, such

as laboratory analyses and field surveys, are often time-consuming, costly, and limited in spatial coverage. This work aims to develop an intelligent system for

predicting soil quality and recommending suitable crops using machine learn- ing and geospatial data. To achieve this, two key experiments were conducted.

In the first experiment, four models (RBFN, LightGBM, XGBoost, and DNN) were applied to SoilGrids data, including physical and chemical characteristics of the soil. The XGBoost model achieved the best performance R

2 = 0.98, re- asserting its suitability for SQP tasks. The second experiment used a two-stage

prediction architecture. The first stage trained 36 separate regressors to pre- dict soil and environmental conditions from geolocation data. These predictions

were then used in a Random Forest model to estimate the Soil Quality Index

(SQI). The second stage employed a cosine similarity-based method to recom- mend the most suitable plant species based on the predicted site conditions.

The entire system was deployed as an interactive web application, where users can query real-time SQI maps and receive personalized crop recommendations. Future enhancements include incorporating more localized data, particularly from Algeria, and expanding the spatial coverage of the system.