COMPARATIVE ANALYSIS OF REGULARIZED LINEAR REGRESSION TECHNIQUES FOR SOIL ORGANIC CARBON PREDICTION IN LARGE-SCALE SOIL DATASETS

Soil Organic Carbon (SOC) is a key indicator of soil fertility, ecosystem sustainability, and agricultural productivity, making its accurate assessment essential for data-driven and digital agricultural systems. In recent years, machine learning approaches have been increasingly applied to SOC prediction; however, the performance and limitations of baseline linear regression models on large-scale routine soil datasets remain insufficiently explored. This study presents a comparative evaluation of linear regression models for SOC prediction using a large soil chemistry dataset collected from routine laboratory analyses in Uzbekistan between 2022 and 2024. The dataset includes spatial coordinates and agrochemical indicators such as P₂O₅ and K₂O, with SOC (humus content) as the target variable. Linear Least Squares, stochastic gradient descent with Adam optimization, and mini-batch gradient descent were implemented and evaluated using standardized inputs. The experimental results demonstrated consistent generalization across all optimization strategies, achieving test-set R² values of approximately 0.31. Although the models exhibit stable behavior and interpretable coefficients, their limited predictive accuracy highlights the inherently nonlinear nature of SOC dynamics. These findings emphasize the role of linear models as transparent baselines and motivate the use of advanced nonlinear machine learning techniques in future digital soil analytics systems.