Synergizing Residual and Dense Architectures for Fine-Grained Oil Palm Grading: A Deep Feature Concatenation Approach

Accurate grading of Oil Palm Fresh Fruit Bunches (FFB) is pivotal for maximizing agricultural yield, yet manual assessment in unstructured environments remains labor-intensive and subjective. While Convolutional Neural Networks (CNNs) offer an automated solution, the conventional strategy of scaling network depth often yields diminishing returns or overfitting on moderately sized datasets. To overcome these limitations, this study proposes the Deep Feature Concatenation (DFC) framework. Rather than deepening a single architecture, this methodology synergizes the spatial hierarchy preservation of ResNet50 with the dense feature-reuse mechanisms of DenseNet121. This fusion creates a composite representation space that captures complementary inductive biases. To ensure computational efficiency, the framework decouples representation learning from inference. Principal Component Analysis (PCA) retains 99% of explained variance while compressing features by 68%. These optimized representations are classified using shallow linear probes. Validated on a single-source dataset expanded to 4000 images (derived from 466 original samples) using a rigorous “Parent–Child” split to prevent data leakage, DFC achieved a peak accuracy of 97.75%. McNemar’s statistical test indicated that this performance outperforms the ResNet50 baseline (<inline-formula><math xmlns="http://www.w3.org/1998/Math/MathML" display="inline"><semantics><mrow><mi>p</mi><mo>=</mo><mn>0.039</mn></mrow></semantics></math></inline-formula>) for SVM classifiers. However, it is critical to note that these results represent a proof of concept based on a limited biological sample size, particularly for rare defect classes. While the model achieved 100% detection accuracy for critical defects within the specific validation set, the high synthetic-to-original ratio necessitates cautious interpretation regarding external validity. This framework provides a practical foundation for future research into high-precision, low-latency grading systems, but multi-center validation on larger, independent datasets is required to confirm broad generalizability across diverse plantation environments.