Comparative analysis of deep learning architecture for object-based crop classification using UAV multispectral imagery

Although crop classification using object-based image analysis (OBIA) has been widely investigated, studies integrating OBIA-derived features with deep learning architectures for high-precision crop identification remain limited. This study evaluates four deep learning architectures: Tabular Convolutional Neural Network (Tab-CNN), Tabular Residual Network (Tab-ResNet), Transformer, and a hybrid CNN-Transformer encoder-decoder, for crop type classification using OBIA-derived feature vectors extracted from UAV multispectral imagery. Image segmentation was performed using eCognition software, and spectral, index, and gray-level co-occurrence matrix (GLCM) features were extracted from the segmented object. The models were implemented in Python (Kaggle Notebook environment) using PyTorch 2.6.0, scikit-learn 1.2.2, and Optuna 4.5.0 for hyperparameter optimization. All models achieved strong classification performance, with accuracies and F1-score above 93%. Tab-ResNet produced the best results (accuracy = 95.14%, weighed F1 = 0.95, macro F1 = 0.90, Cohen's kappa κ = 0.94), followed by the Tab-CNN (accuracy = 93.80%, weighed F1 = 0.94, macro F1 = 0.77, κ = 0.92). The Transformer model showed comparable performance, whereas the hybrid CNN-Transformer architecture exhibited slightly lower accuracy. Feature-importance analysis using SHAP revealed that vegetation indices and GLCM-based texture metrics were the most influential predictors. Overall, the results demonstrate that deep learning architectures trained on OBIA-derived features can accurately classify crop types from UAV multispectral. Future work should evaluate the proposed framework using multi-site and multi-temporal UAV datasets to assess model robustness across varying environmental conditions, crop phenological stages, and agricultural regions.