Advanced deep learning framework for soil texture classification

In soil texture classification, accuracy with interpretability is the key to sustainable agriculture and environmental management. The presented ATFEM (Advanced Triptych Feature Engineering and Modeling framework) framework synergizes handcrafted texture features with learned deep representations through a three-stream architecture: VGG-RTPNet (Residual Texture-Preserving Network based on Visual Geometry Group-16) for texture, ResNet-DANet (Residual Network integrated with Dual Attention Network) for semantics, and Swin-FANet (Shifted Window-based Frequency-Aware Network based on Transformer) for spectral spatial correlation. Subsequently, these branches help in extracting fine-grained structural, dual-attention-enhanced semantic, and spectral-spatial correlation-wise features of soil-image data. To further eliminate redundancy from the feature sets and arrive at the best representation, a Feature Fusion and Selection strategy employing an enhanced hybrid metaheuristic method termed EWJFO (Enhanced Wombat-Jellyfish Feature Optimization) is proposed. It synthesizes the adaptive exploration behavior of Wombat Optimization Algorithm (WOA) with the swift control convergence tempo of the Jellyfish Search Optimizer (JSO) to select the best feature subset. In addition, a new handcrafted descriptor for soil texture image analysis referred as Farthing Ornament of Histogram of Oriented Gradients (F-HOG) has been introduced with adapative. Conventional HOG is burdened with having high-dimensional redundancy and hence suffers from noise sensitivity, F-HOG combines the effect of a Butterworth frequency filter to remove the unwanted high-frequency artifacts and then goes on to perform the statistical selection of the most frequent gradient bins, thus reducing dimensions and retaining quite a bit of the discriminative structural information. The experiments were conducted on a self-built soil texture image dataset consisting of 4,000 labeled images distributed among five texture classes. ATFEM achieved an accuracy of 98.10%, an F1 score of 89.60%, Cohen's kappa rating of 94.80%, and an AUC of 98.10%, outperforming cutting-edge methods such as CatBoost-DNN, GBDT-CNN, and SVC-RF. This work offers an upscalable, explainable, and expressively accurate solution for soil texture mapping in precision agriculture and environmental monitoring.