Performing Department
(N/A)
Non Technical Summary
Evapotranspiration (ET) information is critical for climate-smart irrigation practices, but local crop ET forecasts are currently unavailable for most farmers. The overall objective of this project is to develop and deliver a proof-of-concept framework for globally available short-term crop ET forecasts. These forecasts will be tailored to local conditions using information from farmer-provided smartphone images. The project will build from the free Canopeo app already used worldwide. The new proof-of-concept tool will use Canopeo images and metadata in conjunction with globally available weather forecasts and novel deep-learning models. Weather forecasts will be used to forecast reference ET, which will be adjusted to estimate the ET of the farmer's particular crop (e.g., via a crop coefficient) using an established method informed by the percent canopy cover measured by Canopeo. To further tailor ET estimates to the soil moisture conditions (i.e., estimate the water stress coefficient) is more difficult because there is no establishedscalable method for estimating local soil moisture in irrigated crops. To fill this gap, a novel deep-learning method will be developed. This method will estimate the water stress coefficient based on recent weather and crop (and soil) conditions as reflected in the mobile phone images. The forecasts will be delivered to beta testers and their feedback will be used to refine the crop ET forecast system. Beta testers will come from the 30k+ Canopeo users worldwide and from Oklahoma irrigators engaged through the Oklahoma State University Extension network.
Animal Health Component
0%
Research Effort Categories
Basic
(N/A)
Applied
100%
Developmental
(N/A)
Goals / Objectives
The overall goalof this project is todevelopanddelivera proof-of-concept model for globally available crop ET forecasts based on smartphone images and tailored to local soil moisture conditions using deep learning (DL).To accomplish our overall goal, we will focus on three objectives:Extend existing methods to forecast site-specific crop ET for well-watered conditions anywhere globally. Develop a deep learning model to estimate crop water stress based on recent weather and images provided by the farmer.Deliver the proof-of-concept model and collect feedback.
Project Methods
Reference ET is the estimated ET that would occur from a reference crop (grass or alfalfa) in ideal growing conditions (well-watered and relative humidity above 45%). Reference ET is adjusted based on weather conditions including relative humidity, wind speed, and crop height. The crop coefficient is the ratio between the crop ET and the reference ET and further adjusts the reference ET to reflect the crop-specific water usage at the crop's given growth stage and level of canopy development. Finally, the ET estimate is adjusted by the water stress coefficient. When the crop is water stressed, the crop water stress coefficient drops below 1, and crop ET is reduced.Objective 1:Extend existing methods to forecast site-specific crop ET for well-watered conditions anywhere globally.EffortsWe plan to accomplish the objective for this aim in the following steps. Each of these steps will be prototyped locally first offline then they will be transitioned to a cloud environment for use in beta testing.Develop a data pipeline for retrieving globally available weather forecasts to forecast reference ET.We will use NOAA's global forecasting system (GFS) to obtain weather forecasts needed for forecasting reference ET. GFS forecasts include temperature, wind speed, relative humidity, and cloud cover which are needed to compute reference ET.Implement the A&P method to estimate crop coefficient.To estimate the crop coefficient, we will use the proven and versatile A&P method which models crop coefficient based on percent canopy cover and the crop height. Both percent canopy cover and crop height will be supplied by participating users through the Canopeo app.Combine the reference ET and crop coefficient to produce a site-specific crop ET forecast for well-watered conditions.We will develop and deliver, via automated e-mails, 7-day forecasts as an alpha version of the Crop ET Forecasts. The forecast time window can be adjusted from 1 to 16 days based on user feedback.EvaluationThe efforts of Objective 1 will be evaluated based on the ability of the tool to produce a reference ET forecast and tailor that forecast based on the crop coefficient. We will also evaluate this based on feedback from ET forecast users.Objective 2:Develop a deep learning model to estimate crop water stress based onrecent weather and images provided by the farmer.EffortsIn this objective, we will iteratively develop novel deep-learning frameworks to estimate the soil moisture conditions from recent weather and images provided by farmers in theCanopeomobile app. We propose the followingtwo-foldcontributions in this proposal.Deep Learning Model to Quantify Crop Water Stress in Images: We propose an end-to-end deep learning framework that provides an initial estimate of crop water stress directly from raw images. The images from the Canopeo app often include crop canopy, crop residue, and soil. Thus, the prediction model involves two sub-tasks: (i) Identify soil and crop canopy in raw images, and (ii) Estimate the water stress from these image segments. We propose a multi-scale object detection-based deep learning model [22, 23, 24] calledMap-Soilto identify different object types like plant, soil, residue, shadows, and other unknown objects. Importantly, the proposed model will utilizeResNet, a pre-trained object detection model, as a backbone architecture and it will also include Graph Neural Networks (GNNs) to capture robust global representations of images from their pixels in addition to their local features. OurMap-Soilmodel will identify bounding polygons around image objects like soil and leaves which will be further used for the water stress prediction task. For that task, we will build on the gradient boosting decision tree (GBDT), which was able to distinguish three levels of water stress in maize (full irrigation, deficit irrigation, and drought stress). By first segmenting the soil and crop portions of the images, we expect to be able to provide a more detailed categorization and quantification of the crop water stress. Team members' existing datasets with thousands of soil moisture observations and collocated Canopeo images will be used to train and evaluate the proposed model. We will evaluate our model predictions with machine learning metrics like accuracy, precision, recall, F1, and mean absolute error.Knowledge-Infused Hindcast Model:The existing work, use only the image details to estimate water stress. To advance the state of the science, we propose to refine those estimates via a Knowledge-Infused modelto hindcast soil moisture. This model will, for example, consider temporal information collected from external sources like recent (14-day) local weather patterns, soil characteristics, and potentially past photographs (if available) to hindcast soil moisture. Random Forests, XGBoost regression, and Long Short-term Memory approaches have all shown good potential for soil moisture hindcasts and will be explored here.EvaluationThe efforts of Objective 2 will be evaluated usingthe error in the model that is used to predict soil moisture from crop photographs.Objective 3:Deliver the proof-of-concept model and collect feedbackEffortsWe plan to accomplish the objective for this through the following steps.Deliver forecasts to interested Canopeo users.Starting with its first iteration (Objective 1), we will offer the crop ET forecasts to the existing Canopeo user base. We will ask for their feedback about the usefulness of the forecasts and the forecast delivery. Since the users providing this feedback could be from anywhere on the globe, this feedback will be collected in a scalable way (e.g., an online survey).Deliver to Oklahoma Irrigators via OSU Extension. We will educate and deliver the model to irrigators in Oklahoma through the OSU Extension Master Irrigator class. The content will be delivered by Dr. Sharma and a graduate student. At these meetings, the process of using the tool will be explained and feedback will be gathered.Deliver to rural Oklahoma producers.An undergraduate participant in OSU's Rural Scholar's program will spend 10 weeks during the summer in rural Oklahoma communities working with irrigators and training them to use the crop ET forecast tool. This student will also collect feedback about the forecasts from the producers. OSU Extension educators in each county will help this Scholar connect with local irrigators.EvaluationThe efforts of Objective 3 will be evaluated based on the number of users (both global and from Oklahoma) that use the crop ET forecast tool. We will also evaluate the efforts based on the number of users we are able to collect feedback.