DEEP LEARNING PREDICTIVE MODEL FOR SOIL TEXTURAL ASSESSMENT

The distribution of grain sizes in different soil samples is essential for agriculture and geotechnics, providing high-resolution soil maps crucial for land use planning. Traditional methods for soil texture analysis are reliable but often time-consuming and inconsistent. With that, this study aims to create an efficient predictive model for soil texture classification using deep learning techniques. A dataset of 4,556 images was extensively pre-processed and trained, with a model chosen for validation due to its low MSE value of 1.18. The model's performance, evaluated through Precision, Recall, and F1 Score, showed weighted averages of 88%, 78%, and 74%, respectively, and an overall accuracy of 94.56%. Validation using 456 images revealed high accuracy for Sandy and Clayey Soils but varying results for Loamy and Silty Soils. In Trial 1, the model achieved over 91% accuracy for all soil textures, with 100% accuracy for Sandy Soil. However, Trials 2 and 3 exhibited decreased accuracy for Loamy and Silty Soils, with the lowest accuracies at 61.40% and 65.78%, respectively. These results suggest that while the model is effective for certain soil textures, it requires further refinement and additional diverse training data to consistently match the reliability of traditional methods.