Abstract:
Accurate soil type classification is essential for sustainable agriculture, environmental monitoring, and effective land management. Due to the inherent variability in soil types and the prevalent class imbalance in soil imaging datasets, reliable classification remains a complex challenge. This study combined four unique soil image datasets, yielding 11 distinct soil type classes. The MobileNet model was modified to enhance classification with essential adjustments: replacing ReLU with Leaky ReLU, incorporating an average pooling layer (APL), and integrating a Long Short-Term Memory (LSTM) Network. Since class imbalance was evident in the extracted features, the Synthetic Minority Oversampling Technique (SMOTE) was employed to balanced these features. The MobileNet-LSTM-SMOTE approach demonstrated strong performance, achieving a Kappa score of 98.6 percent and a 5-fold average Kappa score of 97.9 percent. Additionally, testing on an independent dataset confirmed the model's robustness and generalizability. We also conducted a comprehensive performance comparison, including statistical analysis, against five established models-MobileNet, Xception, ResNet50V2, VGG-19, and DenseNet121. The results underscore the efficacy of proposed method in overcoming the challenges of class imbalance and feature variability, offering a reliable solution for soil type classification.
