UAV-based visible-infrared image fusion can offer a promising tool for rapid identification of rice blast resistance (RBR), which is significant to promote rice breeding and ensure global food security. However, effectively combining the benefits of visible and infrared modalities for RBR identification still remains a challenge. This challenge stems from the difficulty of adaptively balancing the contribution of both modalities, and preserving the integrity of local and global representation for decision making. Also, the widely employed multi-scale feature fusion brings extra computational costs. To address the aforementioned limitations, this paper proposes a parallel resnet-transformer and cross-modal attention (PRCA) network for rapid RBR identification. The hybrid architecture processes the local and global feature learning in separate branches, retaining the representation ability of local and global features to maximum extent. Later, we introduce a visible-infrared cross-modal (VICM) attention mechanism to build the integration of visible and infrared modalities, striving to balance the contribution of visible and infrared imagery. Next, we present a lightweight multi-scale feature fusion (LMFF) method to further improve the precision, introducing no extra parameters and low computational costs. Comparison results of field tests on 530 rice plots showed that the PRCA model outperformed other SOTA methods in OA (+1.85%), kappa (+3.22%), F1-score (+0.24%), and recall (+1.38%). Finally, the effectiveness of our method was validated through ablation study. The experimental results indicate that the PRCA method has potential in RBR identification as well as quantitative inversion of other phenotypic traits.

Accurate detection of citrus can be easily affected by adjacent branches and overlapped fruits in natural orchard conditions, where some specific information of citrus might be lost due to the resultant complex occlusion. Traditional deep learning models might result in lower detection accuracy and detection speed when facing occluded targets. To solve this problem, an improved deep learning algorithm based on YOLOv5, named IYOLOv5, was proposed for accurate detection of citrus fruits. An innovative Res-CSPDarknet network was firstly employed to both enhance feature extraction performance and minimize feature loss within the backbone network, which aims to reduce the miss detection rate. Subsequently, the BiFPN module was adopted as the new neck net to enhance the function for extracting deep semantic features. A coordinate attention mechanism module was then introduced into the network’s detection layer. The performance of the proposed model was evaluated on a home-made citrus dataset containing 2000 optical images. The results show that the proposed IYOLOv5 achieved the highest mean average precision (93.5%) and F1-score (95.6%), compared to the traditional deep learning models including Faster R-CNN, CenterNet, YOLOv3, YOLOv5, and YOLOv7. In particular, the proposed IYOLOv5 obtained a decrease of missed detection rate (at least 13.1%) on the specific task of detecting heavily occluded citrus, compared to other models. Therefore, the proposed method could be potentially used as part of the vision system of a picking robot to identify the citrus fruits accurately.