Progress 09/01/18 to 08/31/22
Outputs
Target Audience:The outcomes of this project were used to communicate with scholars, growers, scientists, manufacturing industry, and other stakeholders from around the world through publication and presentation in local, national, and international meetings and conferences. Project activities and outcomes were also discussed with news outlets (e.g., Good Fruit Growers), which disseminated the information to growers, researchers, and public in the Pacific Northwest region as well as nationally. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided? Three Ph.D. students, one master's student and one visiting scholar were actively involved in this project. PIs, students, and scholars interacted frequently (monthly meetings plus additional meetings as desired) to discuss the progress, address challenges, and plan future tasks and activities. Students carried out most of the day-to-day research activities, including data collection and analysis. Students were also supervised for research paper writing, presentation, and publications. The project provided research internships for five undergraduate students to participate in scientific investigations. The interns got the opportunity to work on the project with leading experts. How have the results been disseminated to communities of interest? Several sensor platforms with multiple sensors were developed and used successfully for data collection in Washington State University research vineyards. We presented different aspects of the data collection, integration and analysis methods as well as results in various local, national and international conferences and grower meetings, as listed in section 4. These presentations initiated the dissemination of findings to various stakeholders and initiated a discussion in the scientific community for future research and development. Our research activities were introduced to the communities of interest by demonstrating to and publishing our work in Good Fruit Grower magazine, which is widely distributed to orchardists and vineyardists worldwide. The research findings were introduced to >150 growers and other wine industry stakeholders during the 2021 Viticulture Field Day at IAREC and other field days and grower-organized meetings. What do you plan to do during the next reporting period to accomplish the goals?
Nothing Reported
Impacts
What was accomplished under these goals?
1) Determine crop-specific tools for detecting plant water status Different irrigation treatments or regimes were established in a Riesling vineyard block at the WSU Irrigated Agriculture Research and Extension Center to support the project objectives. These treatments successfully generated two distinct soil environments that translated to measurable differences in common viticultural indicators of canopy growth and plant water status (leaf water potential, yield, shoot number and length, and pruning weight). These differences were statistically significant across three years, providing a realistic and heterogeneous environment to capture the potential spectral and geometric signature differences in field-grown grapevines. Two additional treatments were added in 2021 to generate "extreme" cases by completely withholding irrigation water after fruit set (severe water stress) and by irrigating to field capacity 3 days prior to data collection (no water stress). Overall, the field irrigation trial provided a reliable background for big-data analytics and further research. Intensity of linear relationships between Vegetative Indices (VIs, derived from hyperspectral data) and plant water stress (indicated by leaf water potential, ΨL) and repeated patterns/trends in those relationships were analyzed for the selection of important VIs (Data from 2019 and 2020). An optimized Random Forest Classifier (RFC) model and an Artificial Neural Network (ANN) were developed for the classification of leaf water potential into three levels of water stress. In addition, several weather parameters (relative humidity, evapotranspiration, minimum soil temperature, and solar radiation) demonstrated high significance in model development and were included in the final optimized RFC and ANN models. 2) Develop robust and reliable approaches for estimating plant water status using multi-source data and Big Data analytics A custom-developed algorithm was created to automatically derive canopy temperature (Tc) and calculate crop water stress index (CWSI) from the acquired thermal-RGB images. The relationship between leaf water potential (Ψleaf) and CWSI was investigated. The results revealed that the proposed algorithm combining thermal and RGB images to determine CWSI can be used for assessing crop water status of grapevines. A water stress diagnosis model for grapevine was built by combining the hyperspectral images and leaf 3D point cloud data. 3D point cloud data could reveal the leaf orientation parameters, complementing the critical information in the complex interactions of light and plants. When discriminating the grapevine water status with VNIR hyperspectral imaging, the accuracy was improved by fusing the spectral and 3D point cloud data. Four major components of a big data analytics knowledge base platform were built and enabled. The platform (1) maintains the underlying facts as knowledge from the experimental observations from the field data, which contains leaf water potential and stomatal conductance for all three treatments; (2) integrates an ontology to capture similarity between different treatments and measurements and improve the precision of predictive models; (3) cleans the noisy agriculture sensor data and copes with missing values; and (4) provides a data exploration interface for domain experts to easily navigate and digest the data. Refined the knowledge base platform for smart irrigation analysis with the new data in 2020/2021 incorporated. Enhanced the knowledge base platform with new factual knowledge extracted from the collected sensor data. 3) Develop decision-support tools for automated scheduling of precision irrigation. We developed a decision-support system for managing precision RDI in vineyards. The system consists of a soil moisture prediction model and an RDI scheduling model developed based on artificial neural networks (ANN). Initial soil moisture, weather variables, crop coefficient, and irrigation amount were used as inputs to the soil moisture prediction model. The output from this prediction model provides an indicator of future water status in soil and vines. Initial soil moisture, weather variables, crop coefficient and desired soil moisture target were used as inputs to the RDI scheduling model for regulating the amount of water applied to achieve a desired soil moisture target which is connected to the grapevine water stress target. Developed and evaluated new modules for predictive modeling and analysis based on regression analysis. Given the measurements observed in a week, the model predicts the number of irrigation hours for the next week, which is close to the number provided by human experts. In the evaluation phase, the predictive model achieved an RMSE of 5.5 hours with R2 of 0.86 on Regulated Deficit Irrigation (RDI) treatments. Conducted experiments for the predictive power of contextualized modeling and few-shot learning (FSL) using the 2019 and 2020 data. These methods allow predictive models to be learned with only a small amount of observed data. The contextualized learning and few-shot learning further improved the predictive models with RMSE reduced from 11.8 hours to 5.8 hours and 3.5 hours, respectively, compared to their counterparts trained from the whole dataset collected. Developed a new class of spatiotemporal graph neural network (st-GNN) model for irrigation prediction. We have tested the st-GNN model using spatiotemporal sensor data from other domains and verified its effectiveness. We have also tested its application on detecting and repairing data errors of historical data.
Publications
- Type:
Journal Articles
Status:
Published
Year Published:
2019
Citation:
Lin, P., Song, Q., Wu, Y., & Pi, J. (2019). Discovering Patterns for Fact Checking in Knowledge Graphs. Journal of Data and Information Quality (JDIQ), 11(3), 13.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2020
Citation:
Lin, P., Song, Q., Wu, Y., & Pi, J. (2020). Repairing Erroneous Entities using Star Constraints in Attributed Graphs. In 2020 IEEE 36th International Conference on Data Engineering (ICDE). IEEE.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2021
Citation:
Song, Q., Lin, P., Ma, H., Wu, Y., (2021). Explaining Missing Data in Graphs: A Constraint-based Approach. International Conference on Data Engineering (ICDE).
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2021
Citation:
Guan, S., Ma, H., Choudhury, S., & Wu, Y. (2021). GEDet: detecting erroneous nodes with a few examples. Proceedings of the VLDB Endowment, 14(12), 2875-2878.
- Type:
Theses/Dissertations
Status:
Published
Year Published:
2021
Citation:
Thapa, S. (2021). Irrigation Management in Vineyards: Modeling Water Stress using Hyperspectral Imaging. MS Thesis, Washington State University.
- Type:
Journal Articles
Status:
Published
Year Published:
2022
Citation:
Zhou, Z., Diverres, G., Kang, C., Thapa, S., Karkee, M., Zhang, Q., & Keller, M. (2022). Ground-Based Thermal Imaging for Assessing Crop Water Status in Grapevines over a Growing Season. Agronomy, 12(2), 322.
- Type:
Journal Articles
Status:
Published
Year Published:
2022
Citation:
Thapa, S., Kang, C., Diverres, G., Karkee, M., Zhang, Q., & Keller, M. (2022). Assessment of Water Stress in Vineyards Using On-The-Go Hyperspectral Imaging and Machine Learning Algorithms. Journal of the ASABE, 0.
- Type:
Journal Articles
Status:
Accepted
Year Published:
2022
Citation:
" Kang, C., Diverres, G., Karkee, M., Zhang, Q., & Keller, M. Decision-support System for Precision Regulated Deficit Irrigation Management for Wine Grapes. Computers and Electronics in Agriculture. Accepted.
- Type:
Other
Status:
Published
Year Published:
2020
Citation:
Diverres, G., Keller, M. (2020). Evaluating deficit irrigation strategies in Washington State. Poster and oral presentation, Washington Wine Growers Annual Meeting and 2020 Trade Show, Kennewick, WA.
- Type:
Other
Status:
Published
Year Published:
2021
Citation:
Diverres, G., Keller, M. (2021). Irrigation for white wine grapes. WSU-WSGS Viticulture Field Day. Oral presentation, Irrigated Agriculture Research and Extension Center, Prosser, WA.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2022
Citation:
Kang, C., Karkee, M., Zhang, Q., Keller, M., Diverres, G. (2022). Hyperspectral imaging and stereovision?based water stress assessment in vineyards. Oral presentation, 2022 American Society of Agricultural and Biological Engineers Annual International Meeting, Houston, TX.
|
Progress 09/01/20 to 08/31/21
Outputs
Target Audience:The outcomes of this project were used to communicate with scholars, growers, scientists, the manufacturing industry, and other stakeholders from around the world through publication and presentation in local, national, and international meetings and conferences. Project activities and outcomes were also discussed with news outlets (e.g. Good Fruit Growers), which disseminated the information to growers, researchers and the general public in the Pacific Northwest region as well as nationally. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?3 PhD students and one master's student were actively involved in this project. PIs, students, and scholars interacted frequently (monthly meetings plus additional meetings as desired) to discuss the progress, address challenges, and plan future tasks and activities. Students carried out most of the day-to-day research activities including data collection and analysis. Students were also supervised for research paper writing, presentation, and publications. How have the results been disseminated to communities of interest?A new sensor platform with two more new sensors was developed and used successfully for data collection in Washington State University research vineyards. We presented different aspects of the data collection, integration and analysis methods as well as results in various local, national and international conferences and grower meetings such as "2021 Washington Winegrowers" and "2021 American Society of Agricultural and Biological Engineers Annual International Meeting". These presentations initiated the dissemination of findings to various stakeholders and initiated a discussion in the scientific community for future research and development. Our research activities were introduced to the communities of interest by demonstrating to and publishing our work in Good Fruit Grower magazine, which is widely distributed to orchardists and vineyardists worldwide. What do you plan to do during the next reporting period to accomplish the goals?i) Co-registration of hyperspectral images with acceptable accuracy and development of 3D point-clouds through disparity map generation. ii) Further development of inference analysis framework and adoption of machine learning approaches to enable big data analytics features. iii) Improvement of decision-support tools for automated scheduling of precision irrigation.
Impacts
What was accomplished under these goals?
This study focuses on study of plant water stress using pre-established physiological measurement procedures and an innovative non-contact sensing approach to study patterns within heterogenous data-sources using Big-data analytics. The patterns are expected to help build a decision support system that can help growers to apply optimal amount of water. Obj.1# Determine crop-specific tools for detecting plant water status. 1) Major activities completed / experiments conducted; A field trial involving two irrigation treatments: Full Irrigation (FI) and Regulated Deficit Irrigation (RDI) was initiated in an experimental vineyard. Intensity of linear relationships between Vegetative Indices (VIs, derived from hyperspectral data) and plant water stress (indicated by leaf water potential, ΨL) and repeated patterns/trends in those relationships were analyzed for selection of important Vis (Data from 2019 and 2020). An optimized RFC model and an Artificial Neural Network (ANN) were developed for classification of leaf water potential into three levels of water stress. 22 additional access tubes were installed to monitor soil moisture and characterize spatial variability in the experimental vineyard. 2) Data collected Dataset 1 (2020) A well illuminated, fully expanded middle aged sample leaf per canopy was tagged, and weekly measurement of leaf water potential was made. Hyperspectral and 3D images of entire canopies were acquired. Seasonal weather data were collected from the WSU AgWeatherNet meteorological station. Volumetric soil moisture content was measured weekly at three different soil depths (30, 60, 90 cm) using a neutron probe. Vegetative data to characterize growth and canopy architecture were collected during the season at three key phenological stages. Crop yield and its components (number and weight of fruit clusters, number of berries per cluster, berry weight) were recorded at harvest. Fruit maturity was assessed at harvest by measuring berry juice, total soluble solids, titratable acidity, and pH. Dataset 2 (2021) 40 vines with variable vigor were selected as data vines for data collection. Measurements were taken in three sample leaves in a single canopy. A new hyperspectral sensor (Nano-Hyperspec®) was used to acquire spectrum images. Leaf gas exchange was measured on the same leaves. 3) Summary statistics and discussion of results; It was found that GNDVI (rdataset1 = 0.35; rdataset2 = 0.33) and PRI (rdataset1 = 0.30, rdataset2 = 0.33) had consistent linear relationships with water stress (p < 0.05). It was also found that GNDVI, PRI and ANT contributed approximately 72% to a RFC model involving five VIs (that had good linear correlation with ΨL) as input features. ANT also had the most significant relationship with ΨL in year 2020 (rdataset2 = −0.44). It was found that relative humidity, evapotranspiration, minimum soil temperature and solar radiation had the highest variable importance scores among all weather variables. At harvest in 2020, the average cluster weight and the total yield was significantly lower in RDI compared to the other treatments, but no fruit quality differences were found. 4) Key outcomes or other accomplishments realized. The study investigated trends in relationship between VIs and leaf water potential for optimal selection of VIs. Selected VIs and important weather variables avoided unnecessary redundancies within machine learning models and improved computational speed. Both RFC and ANN models showed promising ability to classify water status into the desired stress groups. RFC model saturated early and showed potential to outperform an ANN model. Obj.2# Develop robust and reliable approaches for estimating plant water status using multi-source data and Big Data analytics 1) Major activities completed / experiments conducted; Refined the knowledge base platform for smart irrigation analysis with the new data in 2020/2021 incorporated. The data integration is ongoing as new 2021 data is collected. Enhanced the knowledge base platform with new factual knowledge extracted from the collected sensor data. Developed and evaluated new active learning modules for predictive modeling and analysis. Given the measurements observed in a week, the model predicts the number of irrigation hours for the next week, with selective queries to ask human experts to refine the predicted answers. Developed a new class of spatiotemporal graph neural network (st-GNN) model for irrigation prediction. We have tested the st-GNN model using spatiotemporal sensor data from other domains and verified its effectiveness. We have also tested its application on detecting and repairing data errors of historical data. Extended our few-shot learning (FSL) techniques to GNN predictive model and verified that the contextualized learning and few-shot learning improve the predictive models with RMSE reduced from 11.8 hours to 5.8 hours and 3.5 hours, respectively, compared to their counterparts. 2) Data collected; Integrated the field measurement data from 2020-2021 into the initial big data analysis platform built with 2019 data, by monitoring the incoming sensor data (e.g., leaf water potential). Applied error detection and data repairing algorithms based on the new GNN model to fix the data errors and impute missing values in the incoming collected data. 3)Summary statistics and discussion of results; Our experimental study verified that graph neural networks (GNNs) serve as a good candidate to predict time-series data over a sensor network that are spatial-temporally correlated. Our experimental study verifies that st-GNNs can further improve the accuracy over its counterpart using simple regression analysis. We further verified that few-shot learning (FSL) that leverages similar data can effectively continue to improve model accuracy for GNN models. For a model trained from three vines, if it predicts the irrigation hour for the fourth vine, it improves R-squared value from 0.93 to 0.99, compared with the model using the data of the fourth vine only. 4)Key outcomes or other accomplishments realized. An updated version of the smart irrigation knowledge base from heterogeneous data has been built with the updated ontology. A contextual GNN-based predictive model has been built to predict proper irrigation plans, which can capture the temporal and spatial context for specific irrigation planning. Obj.3# Develop decision-support tools for automated scheduling of precision irrigation 1) Major activities completed / experiments conducted; A soil moisture prediction model was built using Neural Network to predict the soil moisture in the following week in the vineyard with the data from the years 2017, 2018, 2019 and 2020. The soil moisture in the current week, weather data, crop coefficient and irrigation amount were used as inputs of the model. An irrigation scheduling model was built using Neural Network to predict the irrigation needs in the following week to reach the desired soil moisture target. The soil moisture in the current week, weather data, crop coefficient and soil moisture target for next week were used as inputs of the model. 2)Data collected; Dataset used was similar to what was collected in Obj. #1. In addition, water scheduling tables/spreadsheets created by the human experts were collected. 3) Summary statistics and discussion of results; The soil moisture prediction model can predict the soil moisture in the following week with an R2 value of 0.95, which could assist farmers to know the future water stress trend. The results can be used to precisely schedule irrigation for the following week to bring the soil moisture content to the desired target. 4) Key outcomes or other accomplishments realized. The soil moisture prediction model and irrigation scheduling model could serve as a decision-support system, which could help growers to achieve the desired soil moisture at key growth stages.
Publications
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2021
Citation:
Diverres, G., Keller, M. (2021). Tailoring smart irrigation strategies for white wine grapes in eastern Washington. Poster presentation, 2021 Virtual 72nd ASEV National Conference.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2021
Citation:
Diverres, G., Keller, M. (2021). Irrigation for white wine grapes. Oral presentation, 2021 Washington State University/Washington State Grape Society Viticulture Field Day.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2021
Citation:
Diverres, G., Thapa, S., Kang, C., Karkee, M., Zhang, Q., Keller, M. (2021). Tailoring smart irrigation strategies for white wine grapes in eastern Washington. Poster and oral presentation, Virtual 2021 Washington Winegrowers Convention.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2021
Citation:
Thapa, S., Kang, C., Diverres, G., Karkee, M., Zhang, Q., Keller, M. (2021). Assessment of Water Stress in Vineyards Using on-the-Go Hyperspectral Imaging and Machine Learning Algorithms. Oral presentation, 2021 American Society of Agricultural and Biological Engineers Annual International Meeting.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2021
Citation:
Kang, C., Karkee, M., Zhang, Q., Keller, M., Peters, T., Diverres, G., Thapa, S. (2021). Decision-Support System for Precision Irrigation in Vineyards. Oral presentation, 2021 American Society of Agricultural and Biological Engineers Annual International Meeting.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2021
Citation:
Guan, S., Ma, H., Choudhury, S, Wu, Y., (2021). GEDet: Detecting Errorneous Nodes with A Few Examples. The 47rd International Conference on Very Large Data Bases (VLDB).
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2021
Citation:
Song, Q., Lin, P., Ma, H., Wu, Y., (2021). Explaining Missing Data in Graphs: A Constraint-based Approach. International Conference on Data Engineering (ICDE).
|
Progress 09/01/19 to 08/31/20
Outputs
Target Audience:The outcomes of this project were used to communicate with scholars, growers, scientists, the manufacturing industry, and other stakeholders from around the world through publication and presentation in local, national and international meetings and conferences. Project activities and outcomes were also discussed with local news outlets, which disseminated the information to growers, researchers and the general public in the Pacific Northwest region as well as nationally. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?3 PhD students, one master's student, and one visiting scholar were actively involved in this project. PIs, students and scholars interacted frequently to discuss the progress, address challenges, and plan future tasks and activities. Students and scholars carried out most of the day-to-day research activities including data collection and analysis. Students and scholars were also supervised for research paper writing, presentation, and publications. How have the results been disseminated to communities of interest?Two sensor platforms were developed and used successfully for data collection in Washington State University research vineyards. We presented different aspects of the data collection and integration methods and analysis in various local, national and international conferences such as '2020 Washington Winegrowers Convention and Trade Show' held in WA (March 2-5, 2020) and '2020 IEEE 36th International Conference on Data Engineering (ICDE)'. These presentations initiated the dissemination of findings to various stakeholders and initiated a discussion in the scientific community for future research and development. What do you plan to do during the next reporting period to accomplish the goals? Co-registration of multispectral images with acceptable accuracy and development of 3D point-clouds through disparity map generation. Further analysis of multispectral and hyperspectral data inclusive of all phenological stages captured and dissemination of conclusive results. Further development of inference analysis framework and adoption of machine learning approaches to enable big data analytics features. Imporvement of decision-support tools for automated scheduling of precision irrigation.
Impacts
What was accomplished under these goals?
Summary of Impacts This study focuses on the study of plant water stress using pre-established physiological measurement procedures and an innovative non-contact sensing approach to study patterns within heterogeneous data-sources using Big-data analytics. The patterns are expected to help build a decision support system that can help growers to apply the optimal amount of water. The irrigation decision thus made has a potential to substantially improve fruit quality while reducing associated costs by optimizing the water supply to different parts of vineyards. The work can be extended in the future to other crops such as apples and cherries. Obj.1# Determine crop-specific tools for detecting plant water status. 1) Major activities completed / experiments conducted; Field experiments involving three irrigation treatments were continued: no-water-stress control; regulated deficit irrigation (RDI); and partial rootzone drying (PRD). Eight probes (GreenShield®) were used to continuously measure soil moisture and soil temperature in 2 replicates per treatment. Extent of relationships between Vegetative Indices (VIs) and leaf water potential (ΨL) and repeated patterns/trends in those relationships were analyzed for selection of optimal VIs. A Random Forest Classifier (RFC) model and an Artificial Neural Network (ANN) were developed for classification of leaf water potential into three levels of water stress: low(over − 0.8 MPa), moderate(between - 0.8 MPa to - 1.2 MPa) and high (below - 1.2 MPa). 2) Data collected; Spectral, thermal, color and 3D images of 24 canopies (for control and RDI) were acquired from the east side of the vines just before solar noon. Measurements were taken using two sample leaves within each canopy. Seasonal weather data were collected from the Washington State University AgWeatherNet meteorological station located approximately 400 m away from the experimental site. Continuous soil moisture and soil temperature data obtained via GreenShield® soil probes at 6 different depths: approximately 10, 20, 30, 50, 70 and 90 cm. Well illuminated, fully expanded middle-aged sample leaves were tagged and photographed to subsequently measure leaf water potential using a Scholander pressure chamber (Model 615, PMS Instruments Co., Albany, OR, USA) from June to September. Vegetative data relative to canopy architecture and growth were collected during the season at three key phenological stages: Fruit set, Veraison and Harvest. Crop yield and its components (number of cluster and weight, number of berries and weight) measured at harvest. Fruit maturity was assessed at harvest by measuring berry juice total soluble solids, titratable acidity, and pH. 3) Summary statistics and discussion of results; Assessment of trends in relationships of VIs and ΨL revealed that GNDVI and PRIhad consistent linear relationships with water stress(p < 0.05). A comparison of variable importance from different RFC models showed that GNDVI, PRI and ANT contributed approximately 72% to a model involving five VIs as input features. The study demonstrated GNDVI, PRI and ANT as the optimal VIs. It was found that relative humidity, evapotranspiration, and minimum soil temperature had the highest variable importance scores in comparison toother weather variables. Final RFC and ANN models developed with optimal VIs and weather variables had 73% and 70% classification accuracy, respectively. FULL and PRD treatments showed denser foliage with longer shoots and fewer exposed clusters when compared to RDI treatment. At harvest, the average cluster weight and the total yield was significantly lower in RDI compared to the same with other treatments. 4) Key outcomes or other accomplishments realized. The study investigated trends in the relationship of different vegetative indices (VIs) with leaf water potential for careful selection of optimal VIs. Selection of these VIs and important weather variables avoided unnecessary redundancies within machine learning models and faster computation. Both RFC and ANN models showed promising ability to classify water stress into the desired groups. RFC model saturated early and showed the potential to outperform an ANN model in classifying water stress. Obj.2# Develop robust and reliable approaches for estimating plant water status using multi-source data and Big Data analytics 1) Major activities completed / experiments conducted; Continued the development of the knowledge base platform for smart irrigation analysis with the new data incorporated in 2020. Enhanced the knowledge base platform with new factual knowledge extracted from the collected sensor data. Developed and evaluated new modules for predictive modeling and analysis based on regression analysis. Given the measurements observed in a week, the model predicts the number of irrigation hours for the next week, which is close to the number provided by human experts. Conducted preliminary experiments for the predictive power of the contextualized modeling and few shot learning (FSL) using the 2019 and 2020 data. These methods allow predictive models to be learned with only a small amount of observed data. The contextualized learning and few-shot learning further improved the predictive models with RMSE reduced from 11.8 hours to 5.8 hours and 3.5 hours, respectively, compared to their counterparts trained from the whole dataset collected. 2) Data collected; Integrated the measurement data of 2020 into the initial big data analysis platform built from 2019 data, by monitoring the incoming sensor data (e.g., leaf water potential) weekly. Applied error detection and data repairing algorithms based on data quality rules to fix the data errors and impute missing values in the incoming collected data. 3) Summary statistics and discussion of results; Our experimental study verified that context-level modeling can significantly improve the prediction accuracy for irrigation hours than a single model. (1) The model using data of FULL and RDI treatments achieved RMSE of about 6.2 hours and 5.8 hours respectively, outperforming the model trained with all data of FULL and RDI treatments.The model built from a total of 18 weeks of data performs much worse. This is because the measurements and planned irrigation significantly varies in the spatial-temporal domains and are better modeled by our contextualized, finer-grained models. Our experimental study verified that few-shot learning (FSL) that leverages similar data can effectively improve model accuracy. For a model trained from three vines, if it predicts the irrigation hour for the fourth vine, it improves R2 value from 0.93 to 0.99, compared with the model using the data of the fourth vine only. F Obj.3# Develop decision-support tools for automated scheduling of precision irrigation 1) Major activities completed / experiments conducted; An irrigation prediction model was built using Partial Least Squire Regression (PLSR) to predict weekly water needs in the vineyard with 2020 data. The amount of water applied from the domain expert's report was used as ground-truth to train the model. 2)Data collected; Dataset used was similar to what was collected in Obj. #1. In addition, water scheduling tables/spreadsheets created by human expert were collected. 3) Summary statistics and discussion of results; The prediction model trained with all available datasets including soil moisture, leaf water potential and weather data achieved the best prediction results. Irrigation hours predicted by the predictive model were close to the decision made by domain experts. The predictive model achieved an RMSE of 5.5 hours with R-Square value of 0.77, which was better than the same with 2019 data. 4) Key outcomes or other accomplishments realized. The results so far are preliminary and need further analysis before a particular conclusion can be delivered.
Publications
- Type:
Other
Status:
Published
Year Published:
2020
Citation:
Diverres, G., Thapa, S., Kang, C., Keller, M., Karkee, M., Zhang, Q., (2020). Evaluating deficit irrigation strategies in Washington State. Poster and oral presentation, 2020 Washington Winegrowers Convention and Trade Show. WA, USA.
- Type:
Conference Papers and Presentations
Status:
Accepted
Year Published:
2020
Citation:
Guan, S., Lin, P., Ma, H., & Wu, Y., (2020). GEDet: Adversarially Learned Few-shot Detection of Erroneous Nodes in Graphs. In 2020 IEEE 36th International Conference on Data Engineering (IEEE BigData). IEEE.
|
Progress 09/01/18 to 08/31/19
Outputs
Target Audience:The outcomes of this project were used to communicate with scholars, growers, scientists, manufacturing industry, and other stakeholders from around the world through publication and presentation in local, national and international meetings and conferences. Project activities and outcomes were also demonstrated and discussed with local news outlets, which disseminated the information to growers, researchers and general public in the Pacific Northwest region as well as nationally. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?3 PhD students, one masters student and one visiting scholar were actively involved in this project. PIs, students and scholars interacted frequently to discuss the progresses, address challenges and plan future tasks and activities. Students and scholars carried out most of the day-to-day research activities including data collection and analysis. Students and scholars were also supervised for research paper writing, presentation and publications. How have the results been disseminated to communities of interest?Two sensor platforms were developed and used successfully for data collection in Washington State University research vineyards from two sides of the canopies. We presented different aspects of the data collection and integration methods and analysis in various local, national and international conferences such as '2019 ASABE Annual International Meeting' held in Boston, MA (July 17-10, 2019). These presentations initiated the dissemination of findings to various stakeholders and initiated a lot of discussion in the scientific community for future research and development. What do you plan to do during the next reporting period to accomplish the goals? Co-registration of multispectral images with acceptable accuracy and development of 3D point-clouds through disparity map generation. Analysis of multispectral and hyperspectral data inclusive of all phenological stages captured and dissemination of conclusive results. Development of inference analysis framework and adoption of machine learning approaches to enable big data analytics features. Development of decision-support tools for automated scheduling of precision irrigation.
Impacts
What was accomplished under these goals?
Summary of Impacts: This study focuses on study of plant water stress using pre-established physiological measurement procedures and an innovative non-contact sensing approach to study patterns within heterogenous data-sources using Big-data analytics. The patterns are expected to help build a decision support system that can help growers to apply optimal amount of water. The irrigation decision thus made has a potential to substantially improve fruit quality while reducing associated costs by optimizing the water supply to different parts of vineyards. The work can be extended in the future to other crops such as apples and cherries. Obj.1# Determine crop-specific tools for detecting plant water status. 1)Major activities completed / experiments conducted; A field experiment was initiated involving three irrigation treatments: no-water-stress control; regulated deficit irrigation (RDI); and partial rootzone drying (PRD). Two sensing platforms were developed to collect different types of canopy images. Diurnal and seasonal evolution of leaf water potential and crop thermal indices, and an optimal correlation between leaf water potential and Crop Water Stress Index (CWSI) were obtained using thermal images. Least squares regression models were developed between all combination of vegetative indices (VIs) attainable with the hyperspectral data and physiological parameters. 2)Data collected; Hyperspectral, multispectral, thermal, color and 3D images of 24 canopies (12 canopies each of two treatments -- control and RDI) were acquired at or just before solar noon from the east side of the vineyard and after solar noon from the west side of the vineyard for a span of 2 months. Seasonal weather data were collected simultaneously from the Washington State University AgWeatherNet meteorological station located approximately 400 m away from the experimental site Volumetric soil moisture content was measured weekly across the experimental field at three different soil depths (30, 60, 90 cm) using a neutron probe. Well illuminated, fully expanded middle aged sample leaves were tagged, and weekly measurements of leaf water potential, stomatal conductance and leaf gas exchange were measured alongside sensor data acquisition using a Scholander pressure chamber (Model 615, PMS Instruments Co., Albany, OR, USA) and a portable infrared gas analysis system Crop yield and its components (number of cluster and weight, number of berries and weight) were measured gravimetrically at harvest Fruit maturity was assessed at harvest by measuring berry juice total soluble solids by refractometry, titratable acidity by titration, and pH using a pH meter. 3)Summary statistics and discussion of results; PRD significantly reduced the number of lateral shoots and leaves, as well as the degree of shoot periderm formation compared to the other irrigation regimes. Further study will be needed to validate this result. Diurnal measurements showed a high correlation coefficient, (R2 = 0.86) between thermal indices and leaf water potential. Significant relationships were observed between a few combinations of wavelengths of hyperspectral data and leaf water potential. Hyperspectral data obtained during the solar noon from east side of canopies led to the maximum correlation with leaf water potential when an index formed with wavelengths 821nm and 840nm were used (linear regression, R2 = 0.39, p47 < 0.0001). Measurements taken at and after solar noon from two sides of the canopies had a significant difference between them (paired t-test, t47 = −4.77, p < 0.0005). 4)Key outcomes or other accomplishments realized. Results obtained so far must be viewed as preliminary and do not permit major conclusions or outcomes. Obj.2# Develop robust and reliable approaches for estimating plant water status using multi-source data and Big Data analytics 1)Major activities completed / experiments conducted; Four major components of a big data analytics knowledge base platform were built and enabled. The platform (1) maintains the underlying facts as knowledge from the experimental observations from the coordinate field data, which contains leaf water potential and stomatal conductance for all three treatments; (2) integrates an ontology to capture similarity between different treatments and measurements and improve the precision of predictive models; (3) cleans the noisy agriculture sensor data and copes with missing values; and (4) provides a data exploration interface for domain experts to easily navigate and digest the data. 2)Data collected; An initial fraction of knowledge base has been implemented by integrating measurement data, experimental data and sensor data. The underlying facts of the knowledge base consists of entities and attributes identified from critical sensor data (e.g., leaf water potential, stomatal conductance, soil water status), and significant spatial-temporal correlation among factors (e.g., treatments). An automatic knowledge repairing algorithm was built and prototyped that can adopt data cleaning rules to detect erroneous information of entities during the knowledge fusion process. A data navigation tool is developed with easy-to-use visualization, which facilitates domain experts to explore, clean, and digest the data. 3)Summary statistics and discussion of results; This project objective is still to be realized and results can be discussed only after the knowledge base design is complete and contains necessary heterogenous data. 4)Key outcomes or other accomplishments realized. An agricultural ontology has been developed, and a smart irrigation knowledge base from heterogeneous data is under construction. For knowledge cleaning, a new algorithm and a prototype of knowledge repairing are developed. A knowledge search tool is developed to integrate and clean the knowledge from unstructured and noisy agricultural sensor network, to prepare for rule-based and data-driven prediction and treatment recommendation. Obj.3# Develop decision-support tools for automated scheduling of precision irrigation 1)Major activities completed / experiments conducted; An irrigation prediction model was built using Partial Least Squire Regression (PLSR) to predict weekly water needs in the vineyard. The amount of water applied from the domain expert's report was used as ground-truth to train the model. 2)Data collected; Dataset used was similar to what was collected in Obj. #1. In addition, water scheduling tables/spreadsheets created by human expert were collected. 3)Summary statistics and discussion of results; The prediction model trained with all available datasets including soil moisture, leaf water potential and weather achieved the best prediction results. 4)Key outcomes or other accomplishments realized. The results so far are preliminary and need further analysis before a particular conclusion can be delivered.
Publications
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2019
Citation:
Thapa, S., Kang, C., Bhattarai, U., Karkee, M., Keller, M., Zhang, Q. (2019). Understanding Spatial and Temporal Variability in Water Status Indicators in Vineyards. Presented in 2019 ASABE Annual International Meeting, Boston, USA.
- Type:
Conference Papers and Presentations
Status:
Published
Year Published:
2019
Citation:
Diverres, G., Thapa, S., Kang, C., Zhou, Z., Keller, M., Karkee, M., Zhang, Q., (2019). Irrigation strategies for wine grapes in eastern Washington. Presented in 2019 Washington State Grape Society (WSGS) Annual Meeting, WA, USA.
- Type:
Journal Articles
Status:
Published
Year Published:
2019
Citation:
Lin, P., Song, Q., Wu, Y., & Pi, J. (2019). Discovering Patterns for Fact Checking in Knowledge Graphs. Journal of Data and Information Quality (JDIQ), 11(3), 13.
- Type:
Conference Papers and Presentations
Status:
Accepted
Year Published:
2020
Citation:
Lin, P., Song, Q., Wu, Y., & Pi, J. (2020). Repairing Erroneous Entities using Star Constraints in Attributed Graphs. In 2020 IEEE 36th International Conference on Data Engineering (ICDE). IEEE.
|
|