OrchLoc is a novel fingerprinting-based localization system for orchards using a single LoRa gateway. It achieves tree-level accuracy by extracting Channel State Information (CSI) from eight channels, which is used as a fingerprint. A CSI Generative Model (CGM) is designed to learn the relationship between CSIs and their corresponding locations, enabling efficient database construction and updating. The CGM is pre-trained with fingerprints from a reference area and fine-tuned with data from sensor nodes in new areas. This turbo-training scheme allows for minimal manual intervention and effective database updates. The system integrates a complex-valued FC block as a classifier to process CSI data, enhancing localization accuracy. Experimental results show that OrchLoc achieves high precision and recall, with a localization error of just 1.2 m. Substituting GPS with OrchLoc for robotic navigation in orchards reduces navigation errors by 61.3%. The system is robust to environmental changes, including temperature fluctuations and foliage density variations, and can be adapted to other environments with similar spatial and media homogeneity. The system is implemented using LoRa nodes and a Raspberry Pi 4, with training and inference processes optimized for agricultural applications. The results demonstrate the effectiveness of OrchLoc in achieving tree-level localization with minimal overhead and improved robot navigation accuracy.OrchLoc is a novel fingerprinting-based localization system for orchards using a single LoRa gateway. It achieves tree-level accuracy by extracting Channel State Information (CSI) from eight channels, which is used as a fingerprint. A CSI Generative Model (CGM) is designed to learn the relationship between CSIs and their corresponding locations, enabling efficient database construction and updating. The CGM is pre-trained with fingerprints from a reference area and fine-tuned with data from sensor nodes in new areas. This turbo-training scheme allows for minimal manual intervention and effective database updates. The system integrates a complex-valued FC block as a classifier to process CSI data, enhancing localization accuracy. Experimental results show that OrchLoc achieves high precision and recall, with a localization error of just 1.2 m. Substituting GPS with OrchLoc for robotic navigation in orchards reduces navigation errors by 61.3%. The system is robust to environmental changes, including temperature fluctuations and foliage density variations, and can be adapted to other environments with similar spatial and media homogeneity. The system is implemented using LoRa nodes and a Raspberry Pi 4, with training and inference processes optimized for agricultural applications. The results demonstrate the effectiveness of OrchLoc in achieving tree-level localization with minimal overhead and improved robot navigation accuracy.