Estimating Soil Variability using GIS and Deep Learning

Abstract

This study explores the creative integration of Feature Selection, Geographic Information Systems (GIS), Variable Rate Technology (VRT), and Soil Sensors to improve precision fertilization strategies within arid agricultural environments, with specific reference to regions near Tacna, Arizona. Two datasets were collected to identify nitrate levels in soil. Two datasets, wheat and lettuce yields were examined to determine nitrates correlation to cation exchange capacity, moisture, organic material, and temperature. A Random Forest Regression (RFR) supervised learning model was developed in python to analyze feature importance. For both datasets, the RFR was tested using 40/60%, 60/40% and 80/20% train splits. This was done to incorporate the changing agricultural landscape and difficulties posed by shifting climate systems. The study offers a dynamic data-informed framework for optimal nutrient allocation. The strategic use of GIS, VRT, and soil sensors enables the implementation of real-time flexibility in arid environments where conventional fertilizing methods contain constraints, resulting in significant savings in resource waste and environmental effects. The study highlights observable improvements made possible by this strategy, verifying its effectiveness in raising the precision of fertilizer application. Subtle, identifiable changes in features contribute towards prediction accuracy around different scales.