Spatio-temporal stability of intelligent modeling for weed detection in tomato fields

CONTEXT

Site-specific Weed Management (SSWM) represents a shift towards precision and sustainability in agricultural weed control by applying treatments selectively. Leveraging machine learning (ML) and deep learning (DL), particularly convolutional neural networks (CNNs), enhances weed detection capabilities through automated image analysis. Challenges such as requiring extensive labeled datasets and spatio-temporal variability of weeds remain. Utilizing multi-year datasets provides an effective solution by reducing labor-intensive annotation efforts and improving model generalization across varying conditions.

OBJECTIVE

This study aimed to develop and assess DL models (YOLOv8, YOLOv9, YOLOv10, RT-DETR) for accurate detection of five weed species (Cyperus rotundus L., Echinochloa crus-galli L., Setaria verticillata L., Portulaca oleracea L., Solanum nigrum L.) in tomato fields by integrating multi-year datasets through temporal, spatial, and spatio-temporal analyses.

METHODS

The study was conducted in commercial tomato fields in Santa Amalia, Badajoz, Spain, across the 2021 and 2022 growing seasons. Images were systematically captured using a high-resolution Canon camera following an “M”-shaped trajectory. Experts annotated weed species manually, and images underwent augmentation and segmentation to enhance dataset robustness. The models were trained on distinct temporal (year-based), spatial (field-based), and spatio-temporal (combined) datasets, utilizing pre-trained weights and consistent hyperparameters. Performance was evaluated using Average Precision (AP) and Mean Average Precision (mAP).

RESULTS AND CONCLUSIONS

Applying DL models effectively detected weed species, highlighting the importance of comprehensive datasets for model performance. The temporal analysis, relying solely on previous-year data, achieved a mean average precision (mAP) of 0.7, demonstrating limited efficacy. Using same-year but different-field data, spatial analysis showed improved accuracy (mAP: 0.81–0.89). The spatio-temporal analysis combining datasets from multiple years provided the highest accuracy (mAP up to 0.91), validating the advantages of integrating diverse data across different periods.

SIGNIFICANCE

Integrating multi-year datasets substantially enhances the precision and efficiency of DL-based weed detection models. This methodology minimizes the dependency on labor-intensive annotation processes, accelerates model deployment, and boosts dataset variability, thus fostering more robust and accurate predictive capabilities. This analysis ensures a comprehensive representation of environmental conditions, soil types, and weed emergence patterns, extending the method's applicability beyond the original sampled areas in tomato commercial fields. Moreover, the environmental and agricultural conditions of the sampled area resemble those found in other major tomato-producing areas worldwide, thus supporting the reproducibility of the experiments and the broader applicability of the study's findings.