Performing Department
(N/A)
Non Technical Summary
National Agricultural Statistics Service (NASS) conducts weekly surveys of crop and soil moistureconditions for U.S. cropland and provides coarse-resolution satellite soil moisture andvegetation conditions via a web application Crop-CASMA. However, its coarse-resolutionmaps are unable to capture field/subfield level soil moisture variations. It is urgently needed todevelop field/subfield-level soil moisture maps for NASS and the agricultural community to monitorcrop growth conditions and assess drought or flood impact. This project will establish a partnershipbetween US and Canadian institutes to develop new Machine Learning High-Resolution SoilMoisture (ML-HRSM2.0) products in support of NASS crop monitoring and assessment. In situnetworks, satellite imagery and model-derived weather, soil moisture, terrain, and soil maps will becombined to predict daily soil water content (SWC) and plant available water storage (PAWS) at 100-m at the surface and rootzone since 2016. We will combine ML models with a process model via dataassimilation to develop crop soil moisture condition mapsand NASS weekly soil moisture condition reports and disseminate allmaps over Crop-CASMA for enhancing NASS soil moisture condition monitoring operation andfor free public use. It is expected that using ML-HRSM2.0 product will help NASS and the Agricultural community improve crop condition monitoring, disaster assessment, andoperational decision makings.
Animal Health Component
0%
Research Effort Categories
Basic
40%
Applied
40%
Developmental
20%
Goals / Objectives
This project will develop the partnership between US and Canadian institutes to develop a new Machine Learning High-Resolution Soil Moisture (ML-HRSM 2.0), which will develop improved surface and rootzone soil moisture data productsin Support of USDA NASS Crop Monitoring and Management across the conterminous US (CONUS). We will achieve the project goal with four specific objectives:1) Use Unsupervised Domain Adaptation (UDA) to improve the performance of the ML models in data-limited regions in the CONUS;2) Combine ML models with a physically based process model via a data assimilation framework to predict field-level soil water content (SWC) and plant available water storage (PAWS) across the CONUS;3) Develop crop soil moisture condition maps based on ML-HRSM maps and NASS weekly soil moisture condition survey reports and disseminate all soil moisture maps over NASS' Crop-CASMA for enhancing NASS soil moisture condition monitoring operation and for free public use;4) Conduct case studies using field-level SWC and PAWS maps for drought and other disaster assessment, crop monitoring, and assessing prevented planting and harvest in the US and Canada.
Project Methods
Specific Objective 1: Use UDA to improve the performance of the ML models in data-limitedregions in the CONUS1.1 Data collection and pre-processing:We have collected, compiled, and pre-processed datasets from in situ soil moisture networksand remote sensing platforms during 2016-2021 and will use these data to build ML models forpredicting soil moisture since Jan 1, 2016.1.2 ML model selection:In this project, different ML models will be first evaluated forpredicting soil water content at 0-5 cm, 0-30 cm, and 0-100 cm.To address the domainshift problem and overcome the scarcity of in situ training samples, we proposed using TransferLearning methods based on Unsupervised Domain Adaptation (UDA).Specific Objective 2: Combine ML models with a physically based process model via a dataassimilation framework to estimate field-level SWC and PAWS across the CONUS2.1 Development of a physically based water balance modelThis project willuse a modified 1-Dimensional three-layer "tipping bucket" model for the water balance model.2.2 Data assimilation of ML model and physical model estimatesThe estimates of daily soil water content/storage at the same depths (0-5, 0-50, 0-100 cm)from the ML model and physical model will be combined using a data assimilation approach.2.3 Model evaluation1) Evaluation based on independent NEON soil moisture dataset.2) Evaluation using case studies in selected sites in the US and Canada.Specific Objective 3: Disseminate soil moisture maps on NASS' Crop-CASMA for free accessThe team will integrate the 100-m soil moisture maps (ML-HRSM 2.0) into the Crop-CASMAplatform.Specific Objective 4: Conduct case studies using field-level SWC and PAWS maps for cropmonitoring, assessing drought and prevented planting in the US and CanadaTwo methods will be used to produce the soil moisture maps. First, the UDA-based ML modelswill be directly applied to the study sites in Canada without using local soil moisture datasets.Second, we will incorporate a subset of the local in situ moisture measurements to train the MLmodels to improve the model performance. The randomly withheld soil moisture stations will beused to evaluate both soil moisture models mentioned above. The models to be evaluated includeestimates of soil moisture from two ML models (without and with in situ training data), andphysical water balance and data assimilation models.