Identification of soil texture and color using machine learning algorithms and satellite imagery

Area of Science:

• * Remote Sensing
• * Soil Science
• * Machine Learning

Background:

• * Accurate soil information is crucial for effective land-use planning and precision agriculture.
• * Traditional soil sampling methods can be time-consuming and expensive.
• * Satellite imagery offers a potential for cost-effective and large-scale soil property estimation.

Purpose of the Study:

• * To estimate soil texture (clay, silt, sand) and color (Hue, Value, Chroma) using satellite imagery.
• * To compare the performance of Support Vector Regression (SVR) and Decision Tree Regression (DTR) models for these estimations.
• * To evaluate the utility of various indices derived from MODIS sensor imagery as input variables.

Main Methods:

• * Soil properties (texture, color) were measured in situ.
• * MODIS satellite imagery indices were calculated and used as input variables.
• * Support Vector Regression (SVR) and Decision Tree Regression (DTR) models were employed for prediction.
• * Error metrics (RMSE, AMAPE, MAE, MSE, RPD) were used for model evaluation.

Main Results:

• * The SVR model demonstrated superior performance compared to the DTR model across all evaluated soil properties.
• * Hue was predicted with the highest accuracy (lowest RMSE), while sand content exhibited the highest prediction error.
• * Good agreement was found between measured and SVR-predicted soil texture data.
• * Significant temporal differences were observed in the satellite indices, but not consistently with soil texture variability.

Conclusions:

• * SVR is a highly effective method for estimating soil texture and color from satellite imagery.
• * Satellite-derived soil information can support land-use planning and precision agriculture initiatives.
• * Future research should explore integrating SVR with optimization techniques like genetic algorithms to further enhance prediction accuracy.