Fusing very high-resolution aerial imagery with Sentinel-1 and Sentinel-2 time series improves multi-year orchard type classification

Accurate mapping of perennial orchards is essential for agricultural management but remains challenging due to spatially discontinuous canopies and subtle phenological differences. This study presents an operational, application-oriented framework that integrates annual Very High-Resolution (VHR) aerial imagery (0.2 m) with monthly Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical time series to classify twelve orchard types across Israel from 2018 to 2023. By utilizing a multi-year dataset rather than a single seasonal cycle, we specifically evaluate model generalization and temporal robustness against inter-annual variability in phenology and imaging conditions. The dataset includes 106,413 labeled parcel-year samples from 19,950 parcels. We evaluate three model families: (i) satellite-only, (ii) VHR-only, and (iii) a multimodal mid-fusion deep-learning model (integrating spatial-structural and spectral-temporal modalities), under the same-year and cross-year (leave-one-year-out) scenarios. The satellite stream is modeled with a Lightweight Temporal Attention Encoder (LTAE) to capture seasonal dynamics, while the VHR stream uses a Swin Transformer backbone to extract hierarchical spatial features from 0.2 m imagery. The fused model achieved the highest performance across all settings (same-year OA = 0.890 ± 0.009; cross-year OA = 0.881 ± 0.014), outperforming both satellite-only (0.83) and VHR-only models (same-year OA = 0.846; cross-year OA = 0.795). In the best-performing fused cross-year scenario, class-wise F1-scores ranged from 0.71 for apple to 0.99 for banana. Fusion provided the largest gains in early-stage orchards (ages 1–3), where sparse canopy cover weakens spectral signals, but VHR imagery reveals planting geometry. Per-class analyses show that fusion benefits decrease with increasing mean NDVI, highlighting strong gains for sparse or deciduous orchards and minimal gains for dense evergreens. Overall, the results demonstrate that phenological time series and VHR spatial structure are highly complementary, enabling robust, operational orchard-type mapping and informing cost-aware workflows that selectively prioritize VHR acquisitions for young or low-NDVI orchards.