Generalized Soil Texture Classification Using Systematic Generation of Mixtures Based on USDA Soil Classification Triangle

Soil texture classification is pivotal in diverse fields, such as agriculture, environmental science, geology, civil engineering, and land management. Traditional approaches to soil texture classification often need help with geographic constraints and limited generalization capabilities. This research develops a pioneering approach to soil texture classification by establishing a universally applicable model free from geographical boundaries. The researchers systematically generate mixture data using the united states department of agriculture (USDA) soil classification triangle as a foundational framework, combining pure samples of sand, silt, and clay in varying proportions. The mixture of soil samples was captured using a multispectral sensor equipped with seven bands. The novelty of this work lies in developing the cascaded multiclass convolutional neural network (CNN) that performs hierarchical, stagewise classification. Each stage progressively focuses on a reduced subset of classes, beginning with all n classes and removing the class with the highest accuracy at each classifier model until only one class remains. K-fold validation was employed during training to ensure the robustness and generalizability of the model. This approach refines the classification process, improving its ability to handle the complex distinctions in soil texture. The model’s performance is evaluated using various evaluation metrics. The overall accuracy of the classifier on test data was reported to be 88.8% across various soil types while also suggesting room for improvement in accurately identifying certain soil types. This research advances state-of-the-art soil texture classification and lays the foundation for future studies in this domain.