Conference Agenda

Retrieving crop nitrogen (N) status is crucial for sustainable N management in agriculture. Nowadays, remote sensing and machine learning (ML) have made it feasible to non-destructively estimate crop N status. However, most ML models heavily rely on labeled training samples, potentially compromising their robustness under changing conditions. To address these issues, this study proposed a physics-guided ML approach that integrates physics-based radiative transfer models and empirical ML models. The proposed method was validated using hyperspectral data collected at both leaf and canopy levels. At the leaf level, we employed an ASD LabSpec spectrometer to acquire leaf VIS-NIR-SWIR spectral reflectance, followed by manual measurements of leaf chlorophyll a+b concentration. At the canopy level, VIS-NIR hyperspectral images were captured using a hyperspectral camera mounted on an Unmanned Aerial Vehicle at various study sites. Concurrently, crop N content was destructively sampled. Compared to the standard ML method, the proposed method demonstrated superior performance, particularly when only a small fraction (less than 5%) of the total training samples were available at the leaf level. At the canopy level, the proposed method achieved on average higher accuracy than ML and vegetation indices for leaf N content (RRMSE ranging from 10.08% to 10.84%) and canopy N content (RRMSE ranging from 13.89% to 25.21%) estimation. These findings underscore the benefits of integrating physics-based models and empirical ML models in the proposed approach. This approach offers a promising alternative for accurate estimation of crop N status, with potential applicability to other crops and typical crop traits.

Nitrogen has a major physiological impact on grapevines. Therefore, it is crucial to accurately analyze the nitrogen content of leaves since it directly affects both the productivity of the vineyard and the quality of wine produced.

For this research hyperspectral images of grapevine leaves were taken during two vital growth phases—bloom and veraison—from a ground-based hyperspectral sensor. There were two cultivars Chardonnay and Pinot Noir in two different experiment sites in Oregon state. The study employs a comprehensive approach combining hyperspectral imaging data (spectral features) and chemical analysis (measured nitrogen content) of each grapevine leaf to assess nitrogen content in grapevine leaves. The data is processed using machine learning techniques, specifically ensemble feature selection and partial least square regression feature selection method. Both feature selection methods select a set of features (Wavelength bands) in visible and near infra-red range to assess nitrogen in grapevine leaves.

The results of Chardonnay showed a good predictive performance with a coefficient of determination (R²) value of 0.68. The Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) are 0.28% and 0.22%, respectively. The results of Pinot Noir showed predictive performance with an R² value of 0.54. The RMSE and MAE are 0.24% and 0.19%, respectively. The nitrogen estimation results showed a reliable degree of accuracy that outperformed conventional chemical analysis methods in terms of time, workforce, and non-destructiveness.

The results highlight that hyperspectral imaging can be used to assess the nitrogen content in grapevine leaves and can improve nitrogen control in agriculture.

Central to success in precision agriculture (PA) is the accurate characterization of the spatiotemporal variability of soil information, including physical and hydrological properties and state variables. Researchers have studied soil information using near-surface geophysical methods such as Ground-penetrating Radar (GPR) and Electromagnetic Induction (EMI). Recently, the integrated GPR-EMI approach has gained attention for providing comprehensive insights into soil investigation while overcoming their intrinsic limitations. This review explores the advancements, challenges, and potential applications of the integrated GPR-EMI technique. GPR offers non-invasive, high-resolution imaging of subsurface and soil water content (SWC), facilitating precise delineation of soil stratigraphy and hydrological features. EMI exploits subsurface soil electrical conductivity variations to infer properties and state variables like soil texture, SWC, and salinity. Integrating these two leads to interactive benefits, enabling a comprehensive understanding of the variability and dynamics inherent in soil. Over the past several years, field experiments were conducted with an integrated GPR-EMI technique for predicting different soil information. Findings reveal that GPR has the potential to calibrate EMI for estimating SWC covering similar sampling volumes. Furthermore, the integrated technique improves the prediction accuracy of soil bulk density non-destructively in the studied site. Integrated GPR-EMI opens challenges like requirements for high-quality data, detailed data processing, and site-specific calibration requirements. Future research should focus on refining integration techniques by incorporating geophysics with petrophysical models, standardizing protocols, advancing data analysis using artificial intelligence and validating findings across diverse agricultural settings to realize the full potential of this integration to support precision agriculture.

This research explores the application of Detectron2 for detecting Colorado potato beetles in high-resolution images from a moving vehicle in agricultural fields. The study addresses challenges like the small size of beetles and distortion from resizing images, which can make them nearly invisible, compounded by the variable backgrounds and lighting of outdoor settings. To mitigate these issues, we implemented an image preprocessing strategy to boost detection performance which depends on cropping 1920x1080 to 640x640 to maintain beetle representation accuracy. Our images were taken using an imaging system mounted on a sprayer boom at 500 mm above the canopy - resulting in a dataset of 331 training and 41 testing images, all while exposing the system to diverse field conditions. We trained two models: one on original resolution images and another on cropped images. Testing showed the cropped image model improved detection rates from 48% to 70%, and F1-Score from 62% to 82%. To further enhance detection accuracy, a transfer learning approach was adopted to overcome the small size of the original dataset. Rather than utilizing a pre-trained model with the open-source images, a model was trained on a dataset of static, 4k images captured through a handheld device, cropped to 640x640. This new model was then fine-tuned with the original, moving vehicle dataset (640x640), and detection accuracy was increased to 85%, with an F1-Score of 92%. Our results offer a promising path forward for spot-spraying technologies, minimizing agrochemical usage while managing pests, in line with the principles of precision agriculture.