Integration DeepLabv3+ applied to RGB images and vegetation indices for nitrogen status in cereal–legume intercropping system

• A novel deep learning approach in agroecological systems • Developed an RGB-based workflow integrating segmentation and vegetation indices • Applied DeepLabV3+ for crop-specific segmentation in intercropping systems • Enabled nitrogen monitoring using only low-cost RGB imagery This study presents a low-cost, non-invasive approach to monitor the nitrogen status of Triticale in a Triticale–Faba bean intercropping system, an agroecological strategy to avoid chemical nitrogen inputs, using consumer-grade smartphone RGB images and deep learning. Three smartphones (Samsung Galaxy A12, Xiaomi Redmi Note 4, and Redmi Note 11) were used to capture canopy images. A DeepLabV3+ model with a ResNet-50 backbone was trained to semantically segment Triticale pixels from mixed canopies. Training the model on patch-based image subsets, rather than full images, substantially enhanced segmentation accuracy (mIoU = 90.64%). The normalized Dark Green Color Index (nDGCI) derived from segmented images at the canopy scale, was evaluated as a proxy for nitrogen status against normalized SPAD (nSPAD) measurements, a tedious leaf scale method. Strong linear relationships were observed between nDGCI and nSPAD (pooled correlation across optical devices of R² ≈ 0.60 and optical device-specific ranging from R² = 0.69 to 0.87). Statistical analyses highlighted significant effects of cropping modality, phenological stage, and device on both indices, yet the method reliably distinguished nitrogen treatments. Device-specific calibration effectively corrected offsets, validating the feasibility of smartphone-based AI for detailed monitoring in intercropped systems. This approach offers a practical cost-effective alternative to conventional tools enabling precision agriculture in agroecological contexts.