Towards CNN-based classification of crop independent agricultural sub-classes

Extensive UAV-based optical imagery collected over agricultural areas — originating from both scientific and non-scientific sources over the past decade and a half—provides valuable support for improving space-borne crop type classification models through transfer learning. This study evaluates the effectiveness of a Convolutional Neural Network (CNN) for classifying economically significant, small-scale agricultural patterns using high-resolution multi-spectral UAV imagery (0.060/0.080 m spatial resolution) across major crop types in Germany. The imagery was collected from three study areas differing in topography, field size, and management practices, targeting fields of Winter Wheat, Spring Barley, Rapeseed, and Corn at phenological stages where crop sub-classes are most distinguishable. The same CNN architecture was consistently applied and trained separately for each crop to classify critical sub-classes, such as dry/vital and dry/vital lodged crop, flowering crop, ripening crop, weed infestation, and bare soil. High validation performance was observed across several sub-classes, including bare soil (F1: 0.70-0.93; accuracy: 0.88-0.96), disturbed crop (dry/vital lodged) sub-classes (F1: 0.61-0.94; accuracy: 0.89-0.98), flowering crop (F1: 0.51-0.98; accuracy: 0.81-0.99), ripening crop (F1: 0.91; accuracy: 0.94) and vital crop (F1: 0.21-0.96; accuracy: 0.76-0.98). In contrast, lower performance was recorded for dry crop (F1: 0.50-0.56; accuracy: 0.81-0.87) and weed infestation (F1: 0.00-0.43; accuracy: 0.79-0.94). In the transferability phase, performance remained moderate for bare soil (F1: 0.54-0.70; accuracy: 0.77-0.85), disturbed crop (dry lodged) sub-class (F1: 0.63; accuracy: 0.84), but declined significantly for undisturbed crop sub-classes such as dry crop (F1: 0.30-0.99; accuracy: 0.80-1.0), flowering crop (F1: 0.00-0.82; accuracy: 0.76-0.90), vital (F1: 0.00-0.54, accuracy: 0.27-0.79), and ripening crop (F1: 0.00; accuracy: 0.50), as well as for weed infestation (F1: 0.00; accuracy: 0.96). The inability to generalize these finer subdivisions suggests that excessive granularity of the undisturbed crop sub-classes can reduce model robustness across regions. Notably, Rapeseed demonstrated the strongest transferability across sub-classes (F1: 0.54-0.99; accuracy: 0.79-1.00), indicating more stable generalization, possibly due to greater intra-class consistency and domain similarity. The results highlight the potential for developing unified model or applying transfer learning to support practical applications, such as small-scale yield loss assessment. The use of crop-independent sub-class definitions facilitates broader model generalization, enabling advanced agricultural monitoring across platforms and scales. Furthermore, this work contributes to ongoing efforts in integrating space-borne Synthetic Aperture Radar (SAR) data for agricultural monitoring and supports multi-sensor, multi-scale analysis. • The potential for a single CNN model for crop sub-class classification is explored. • Uses thematic sub-classes for scalable, crop-independent modeling. • CNN models achieved high accuracy in validation but varied on unseen data. • CNNs enable crop-independent sub-class separation on limited, constrained datasets. • CNNs showed robustness to mild lighting variations in UAV imagery.