Machine Learning-Based Estimation of Tractor Performance in Tillage Operations Using Soil Physical Properties

Shaogang Gong, Seung-Min Baek, Seung-Yun Baek, Yong‐Joo Kim, Wan-Soo KimAgronomy 2025 · DOI

TL;DR

This paper introduces machine learning models to predict tractor performance (engine torque, power, slip ratio, axle power) using soil physical properties like moisture, cone index, and particle composition. The study demonstrates that incorporating multiple soil variables significantly improves prediction accuracy, with the CatBoost algorithm showing superior performance, paving the way for more efficient precision agriculture.

What Problem Does This Paper Address?

Tractors are crucial for modern agriculture, and understanding their performance is key to optimizing efficiency and fuel consumption. However, tractor performance is heavily influenced by soil conditions, which are highly variable and complex. Accurately estimating this performance in real-time, especially under diverse soil types and moisture levels, is challenging. Traditional methods often rely on extensive physical testing or simplified models that struggle to capture the intricate, non-linear relationships between soil properties and tractor outputs. This gap limits the ability to implement precision agriculture strategies that adapt to real-time field conditions.

Methods

The researchers employed three machine learning algorithms: Decision Tree (DT), CatBoost, and LightGBM. These algorithms were chosen for their ability to model complex, non-linear relationships. They developed four models (A, B, C, and D) using different combinations of input variables:

  • Model A: Soil Moisture Content (SMC) only.
  • Model B: Cone Index (CI) only.
  • Model C: SMC and CI.
  • Model D: SMC, CI, and Sand Proportion.

The models were trained and validated using field data. Performance was evaluated using metrics like R², Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Relative Deviation (RD). Overfitting was assessed using R² Gap and Train-to-Test Loss Ratio (TTLR). SHAP (SHapley Additive exPlanations) was used for sensitivity analysis to understand the feature importance.

Datasets

The study utilized field data collected from 10 paddy fields in Chungcheongnam-do and Gyeonggi-do, Republic of Korea. A total of 600 data samples were collected. The soil physical properties measured included:

  • Soil Moisture Content (SMC)
  • Cone Index (CI)
  • Sand, Silt, and Clay proportions

The tractor performance indicators measured were:

  • Engine Torque (ET)
  • Engine Power (EP)
  • Slip Ratio (SR)
  • Axle Power (AP)

Key Insights and Ideas

  • Soil matters, a lot: Soil physical properties are not just environmental factors; they are critical drivers of tractor performance. Minor changes can have significant impacts.
  • More data, better predictions: Simply put, using more relevant soil variables (SMC, CI, and sand proportion) leads to substantially better tractor performance predictions than using just one or two.
  • CatBoost reigns supreme: For this complex problem, the CatBoost algorithm consistently outperformed DT and LightGBM, especially when using a comprehensive set of soil inputs. Its ability to handle non-linearities and interactions is a key advantage.
  • SMC is king for torque and power: Soil moisture content (SMC) emerged as the most influential factor for predicting engine torque (ET) and engine power (EP).
  • Sand plays a crucial role in axle power: While SMC is dominant for ET and EP, sand proportion becomes the most critical factor for predicting axle power (AP). This highlights how different performance metrics are influenced by distinct soil characteristics.

Key Results

  • Model D (SMC, CI, Sand) outperformed others: This model achieved the highest R² values across all performance indicators, with ET reaching an R² of 0.990 with CatBoost.
  • CatBoost's superiority: CatBoost-based Model D achieved R² values 7.0–14.2% higher than DT and 1.6–3.8% higher than LightGBM for ET, EP, SR, and AP.
  • Significant improvement with multiple variables: Models using multiple soil variables (C and D) showed significantly better performance than univariate models (A and B). For instance, Model B (CI only) had very low R² values, especially for SR and AP.
  • Low overfitting risk: The study found minimal evidence of overfitting, with R² Gap and TTLR values generally within acceptable ranges, indicating good generalization performance.

Limitations and Future Work

  • Limited data scope: The study was conducted in specific paddy field conditions in the Republic of Korea. Generalizing these findings to diverse soil types, climates, and operating conditions requires further validation.
  • Computational cost: While more variables improve accuracy, they also increase computational complexity. Model C, using fewer, real-time measurable variables, showed promising performance, suggesting a trade-off between accuracy and practicality.
  • Future directions: The authors suggest expanding data collection to include a wider range of soil types and operating conditions. Further research into hyperparameter optimization and regularization techniques is recommended to further reduce overfitting and enhance model robustness for real-world agricultural applications.

Summary updated

Abstract

Accurate estimation of tractor performance under various soil conditions is essential for enhancing operational efficiency in precision agriculture. This study developed machine learning models to estimate tractor performance based on key soil physical properties. Three algorithms—decision tree (DT), CatBoost, and LightGBM—were employed to capture nonlinear relationships between soil parameters and tractor performance indicators. The input variables included soil moisture content, cone index, and particle composition, while the output variables were engine torque, power, slip ratio, and axle power. The models in this study were trained and validated using field data collected from eight paddy fields in Chungcheongnam-do (two in Seosan, two in Cheongyang, and four in Dangjin) and two paddy fields in Gyeonggi-do (Anseong), Republic of Korea. Results showed that models using multiple soil variables significantly outperformed those using single variables. In Model D, CatBoost demonstrated superior performance in predicting engine torque, engine power, slip ratio, and axle power, achieving R2 values that were 7.0–14.2% higher than those of DT and 1.6–3.8% higher than those of LightGBM. These findings demonstrate the feasibility of using machine learning with minimal input data to estimate tractor performance, potentially reducing the reliance on extensive physical testing.
BibTeX
@inproceedings{gong2025machine,
  author = {Shaogang Gong and Seung-Min Baek and Seung-Yun Baek and Yong‐Joo Kim and Wan-Soo Kim},
  title = {{Machine Learning-Based Estimation of Tractor Performance in Tillage Operations Using Soil Physical Properties}},
  booktitle = {Agronomy},
  year = {2025},
  doi = {10.3390/agronomy15092228},
}
Figure