Source: GEORGE MASON UNIVERSITY submitted to
FACT: MACHINE-LEARNING-BASED IN-SEASON CROP MAPPING AND ASSOCIATED CLOUDBASED BIGDATA CYBERINFRASTRUCTURE TO SUPPORT USDA NASS DECISION MAKING
Sponsoring Institution
National Institute of Food and Agriculture
Project Status
EXTENDED
Funding Source
Reporting Frequency
Annual
Accession No.
1025609
Grant No.
2021-67021-34151
Project No.
VA.W-2020-08826
Proposal No.
2020-08826
Multistate No.
(N/A)
Program Code
A1541
Project Start Date
Feb 25, 2021
Project End Date
Feb 24, 2025
Grant Year
2021
Project Director
Di, L.
Recipient Organization
GEORGE MASON UNIVERSITY
4400 UNIVERSITY DRIVE
FAIRFAX,VA 22030
Performing Department
(N/A)
Non Technical Summary
The National Agricultural Statistics Service (NASS) of U.S. Department of Agriculture produces annually a digital product, called Cropland Data Layer (CDL), covering the contigious U.S. since 2008. The product maps what crop grows in each field at ~95% accuracy for major crops. CDL is distributed to 60,000 decision makers via CropScape web service system and is one of the key data products for many agricultural decision makings. However, currently there are two major problems associated with CDL production and distribution: 1) the current-year CDL is not available to public until Feb/March next year, making many in-season agricultural decisions impossible; and 2) CropScape on traditional servers cannot meet the peak user requests, especially when a new CDL is released. This project will solve the two problems by 1) producing pre- and in-season CDL-like products through smart algorithms that can learn from historical CDL products and the current satellite observations of the ground; 2) enhancing CropScape to make the new products and the smart algorithms available and easily usable by decision makers. The project will enhance agricultural decision making by providing timely and accurate in-season CDL-like products to NASS and 60,000 unique decision makers through the CropScape cyberinfrastructure.
Animal Health Component
0%
Research Effort Categories
Basic
30%
Applied
35%
Developmental
35%
Classification

Knowledge Area (KA)Subject of Investigation (SOI)Field of Science (FOS)Percent
9030120303050%
9010120209030%
4020120208020%
Goals / Objectives
The goal of the project is to facilitate timely and informed agricultural decision making by developing the capability of generating and providing in-season CDL-like crop maps for CONUS through easy-to-use cyberinfrastructure. The specific objectives of this project is to 1) develop bigdata classification algorithms to automatically derive in-season CDL for CONUS; 2) enhance CropScape by implementing the algorithms as web services; and 3) migrate the enhanced CropScape to a cloud for better user support. In-season CDL means to have CDL-compatible product with reasonable accuracy at beginning of a growing season, continue to improve the product with season progress, and reach the accuracy similar to NASS CDL around early July.
Project Methods
The project will use historical USDA/NASS Cropland Data Layer (CDL) data and current satellite observations to automatically classify and generate pre- and in-season crop planting classification maps of CONUS that has the same classification schema as the CDL. The major satellite observations will be Landsat 8 data. In order to obtain the complete coverage of CONUS and sufficient repetitions over time, other data sources and observations will be considered as well, particularly Sentinel-2 and Sentinel-1 data. Four major activities will be carried out to accomplish the overall goal of producing the pre- and in-season cropland classification from satellite observations and historical CDL data. They are (1) extraction of "trusted pixel" cropland predictions from over 10-year annual historic CDL as ground truth, (2) pre- and in-season cropland classification from satellite observations with optimal deep-learning algorithms, (3) cloud computing for supporting deep-learning classification, and (4) user-experience-centric distribution of cropland classification products using standard geospatial Web services and Web applications on cloud.

Progress 02/25/22 to 02/24/23

Outputs
Target Audience:The project's target audiences are the agricultural sectors, governments, universities, companies, and public and private research teams who demand in-season and pre-season crop information for the United States. Changes/Problems: Nothing Reported What opportunities for training and professional development has the project provided?The project has provided training and professional development in agro-geoinformatics, GIS, remote sensing, crop mapping, land use land cover change, software development, and geospatial cyberinfrastructure for postdoctoral research fellows and Ph.D. students. Specifically, the following opportunities are provided in this reporting period: Biweekly group meetings with mentors, researchers, and students to discuss the progress, methods, data, and tools for the project. Internships for undergraduate students to involve in the project. A seminar entitled "Exploring agro-geoinformation for the United States using the FAIR cyberinformatics tools" was organized at IEEE Geoscience & Remote Sensing Society (GRSS) Washington, DC & Northern VA Chapter Virtual Seminar. Presentations about the latest progress and results of the project were given at International Conference on Agro-geoinformatics. The research findings and insights obtained from the project were incorporated into several graduate and undergraduate courses, such as GGS 773 Interoperability of Geographic Information Systems and GGS 366 Spatial Computing at George Mason University. How have the results been disseminated to communities of interest?The results of in-season and pre-season crop cover maps have been disseminated through the iCrop data service system as well as ArcGIS Server. The iCrop system is a project version of CropScape, the award-winning data service system delivering CDL data products of USDA NASS. According to Google Analytics as of 2022, the CDL and its derivative data products (e.g., crop frequency layers, crop mask layer) have been distributed to over 290,000 users worldwide via CropScape as well as iCrop. With a similar user interface and functionalities, the current user community can smoothly transfer from CropScape to iCrop. Meanwhile, this easy-to-use system will disseminate the project results to new users who demand in-season and crop information for CONUS. The geospatial data on iCrop are disseminated through the OGC web service standard interfaces. When implementing the service layer of iCrop, we used MapServer as the back-end server to power the web mapping capabilities via standardized OGC specifications, such as WMS and WCS. Meanwhile, we used ArcGIS as an alternative back-end WMS/WCS server to host the new geospatial data products described in this study through ArcGIS REST APIs. These standard interfaces enabled interoperability with the front-end client as well as other WMS/WCS-compliant GIS software and applications. What do you plan to do during the next reporting period to accomplish the goals?In the next reporting period, the project will accomplish the following specific objectives: Optimize the in-season mapping algorithms using the upcoming 2022 CDL data. Implement the result distribution service system and integrate them into iCrop. Testing the integrated system. Operational applying the system in producing and servicing pre-/in-season cropland classification data. Release documentation of the system and open source code. Produce the 2023 pre-season crop cover map for CONUS before the growing season (by April 2023) and the in-season crop cover map for CONUS within the growing season (May - July 2023). Write and publish papers about the algorithm, method, data, and system from the project in international journals.

Impacts
What was accomplished under these goals? The goal of the project is to facilitate timely and informed agricultural decision making by developing the capability of generating and providing in-season CDL-like crop maps for CONUS through easy-to-use cyberinfrastructure. The following objectives under these goals have been accomplished in this reporting period: Improve the crop type classification algorithm. The automatic crop type classification algorithm for in-season and pre-season crop cover mapping has been optimized and used for the production of 2022 crop cover maps. The machine learning algorithms, selections of parameters, and prediction models are comparatively studied. The refined algorithm was applied to classify not only dominant crop types (e.g., corn, soybeans, wheat, cotton) but also other minor crop types across the entire United States. Enhance the cyberinformatics tool and system. The web application and services of the cyberinformatics tool designed for this project, iCrop (https://cloud.csiss.gmu.edu/icrop/), were enhanced. The Findable, Interoperable, Accessible, Reusable (FAIR) capability for mapping and geoprocessing has been tested using common GIS clients and software. Specifically, we tested the interoperability of iCrop on the GeoPlatform (https://www.geoplatform.gov/) developed by Federal Geographic Data Committee (FGDC). Survey ground truth for result validation. A set of ground truth data across the United States during the growing season of 2022 was collected and processed. The crop types of over 4,000 land units were surveyed in the DC-Maryland area, Corn Belt area, and Central Valley of California. This ground truth data set was used to validate the 2022 in-season crop mapping results and improve the classification model. Produce and release in-season crop-specific land cover data for 2022. The in-season crop-specific land cover data for 2022 has been produced. The data was created using Landsat 8/9 and Sentinel-2 images acquired within the growing season of 2022. The major crop types of the result can reach 85% - 95% agreement with the ground truth data. The data has been published through the FAIR web services, such as Web Map Service (WMS) and Web Coverage Service (WCS), and released through iCrop as early as August 2022.

Publications

  • Type: Journal Articles Status: Published Year Published: 2022 Citation: Zhang, C., Di, L., Lin, L., Li, H., Guo, L., Yang, Z., Eugene, G.Y., Di, Y. and Yang, A., 2022. Towards automation of in-season crop type mapping using spatiotemporal crop information and remote sensing data. Agricultural Systems, 201, 103462.
  • Type: Journal Articles Status: Published Year Published: 2022 Citation: Lin, L., Di, L., Zhang, C., Guo, L., Di, Y., Li, H. and Yang, A., 2022. Validation and refinement of cropland data layer using a spatial-temporal decision tree algorithm. Scientific Data, 9(1), 63.
  • Type: Journal Articles Status: Under Review Year Published: 2023 Citation: Zhang, C., Di, L., Lin, L., Li, H., Yang, A., Guo, L., Eugene, G.Y., and Yang, Z. iCrop: a FAIR cyberinformatics tool for machine-learning-based U.S. crop-specific land cover data sharing and analysis. Computer and Electronics in Agriculture(Under review)
  • Type: Journal Articles Status: Under Review Year Published: 2023 Citation: Li, H., Di, L., Zhang, C., Lin, L., Guo, L., Eugene, G.Y., Yang, Z. Automated in-season crop-type data layer mapping without ground truth for the Conterminous United States based on multisource satellite imagery. IEEE JSTARS (Under review)
  • Type: Conference Papers and Presentations Status: Published Year Published: 2022 Citation: Zhang, C. Exploring agro-geoinformation for the United States using the FAIR cyberinformatics tools. In IEEE Geoscience & Remote Sensing Society (GRSS) Washington, DC & Northern VA Chapter Virtual Seminar, Online, December 9, 2022.
  • Type: Conference Papers and Presentations Status: Published Year Published: 2022 Citation: Zhang, C. Agriculture 4.0 Technologies for In-season and Early-season Crop Cover Mapping. In 10th International Conference on Agro-Geoinformatics and 43rd Canadian Symposium on Remote Sensing, Session: Agriculture, Online, July 13, 2022.
  • Type: Conference Papers and Presentations Status: Published Year Published: 2022 Citation: Lin, L. Validation and Evaluation of Refined Cropland Data Layer: A Case Study. In 10th International Conference on Agro-Geoinformatics and 43rd Canadian Symposium on Remote Sensing, Session: Agriculture, Online, July 13, 2022.
  • Type: Conference Papers and Presentations Status: Published Year Published: 2022 Citation: Li, H., Di, L., Zhang, C., Lin, L. and Guo, L., 2022, July. Improvement Of In-season Crop Mapping For Illinois Cropland Using Multiple Machine Learning Classifiers. In 2022 10th International Conference on Agro-geoinformatics (Agro-Geoinformatics). IEEE.


Progress 02/25/21 to 02/24/22

Outputs
Target Audience:The project's target audiences are the agricultural sectors, governments, universities, companies, and public and private research teams who demand in-season and pre-season crop information of the United States. Changes/Problems:There were no major changes and problems during this report period. What opportunities for training and professional development has the project provided?The project has provided graduate research assistants, undergraduate interns, and jounir researchers with various training and professional development opportunities, such as one-on-one work with a mentor, internship, study group, and individual study. Specifically, the project initiated a biweekly group discussion to discuss the state-of-the-art methods, data, technologies, and tools for crop mapping. In addition, students and jounir researchers have been given opportunities to present the preliminary result and findings of the project in international conferences and workshops, such as IGARSS, International Conference on Agro-geoinformatics, and AGU Meeting. How have the results been disseminated to communities of interest?The results of in-season and pre-season crop cover maps have been disseminated through the iCrop agro-geoinformatic web service system. The iCrop system is a project version of CropScape, the award-winning web service system delivering CDL data products of USDA NASS. According to Google Analytics as of 2021, the CDL and its derivative data products (e.g., crop frequency layers, crop mask layer) have been distributed to over 250,000 users worldwide via CropScape. With the similar user interface and functionalities, the current user community can smoothly transfer from CropScape to iCrop. Meanwhile, this easy-to-use system will disseminate the project results to new users who demand in-season and pre-season crop information for CONUS What do you plan to do during the next reporting period to accomplish the goals? Validate the CONUS in-season mapping algorithms and results using the 2021 CDL data. Comparatively study alternative learning algorithms, selections of parameters, optimizing the pre-season prediction model. Comparatively study alternative deep learning algorithms, revise the cropland classification algorithm, develop the ensemble algorithm. Produce the 2022 pre-season crop cover map for CONUS before the 2022 growing season (by April) and validate the product with field data and forthcoming 2022 USDA CDL to further validate the algorithm and automatic workflow for pre-season prediction Produce in-season crop cover map for CONUS within the 2022 growing season (May - July), and validate the product with field data and forthcoming 2022 USDA CDL to further validate the algorithm and automatic workflow for in-season mapping Enhance the iCrop system by deploying more web services and data products into it. Write and publish papers about the algorithm, method, data, and system from the project in international journals.

Impacts
What was accomplished under these goals? Automatic crop type classification algorithm prototypes for both in-season crop mapping and pre-season crop cover prediction have been developed and are being evaluated. The algorithms used the "trusted pixels" extracted from the past spatio-temporal crop information to label training samples on satellite images (e.g., Sentinel-2 data) for supervised crop type classification. According to the experiment, it is found that the "trusted pixels" are widely distributed across the CONUS. The prototypical implementation of the algorithm is currently available through Google Earth Engine: https://czhang11.users.earthengine.app/view/agkit4ee-inseason. A CDL-like CONUS in-season crop cover map of 2021 has been produced with the automatic mapping algorithm prototype. The map was created in early August 2021 for the growing season of 2021 by using Sentinel-2 images acquired from April to July. The in-season map is being evaluated with 2021 USDA CDL data, which was released at the end of Febuary 2022. The preliminary evaluation results indicated that the in-season crop map created by this project has reached 85%-95% agreement with the 2021 USDA CDL. The mapping result has been published at iCrop web service portal (described at item 4) through the Web Map Service and Web Coverage Service, which can be used in common GIS clients and software, such as ArcGIS and QGIS. It has been found that the USDA CDL datasets contain some classification errors mainly at the edge of classes, along the road, and at the middle of fields. An algorithm have been developed to correct those errors based on historical CDL time series. Currently 2017-2021 CDL datasets have been refined and published at iCrop. A paper about the algorithm and the refined CDL was published at a prestigious journal, Nature - Scientific Data. An agro-geoinformatic web service portal, iCrop - in-season crop mapping explorer (https://cloud.csiss.gmu.edu/icrop/), has been prototyped and deployed for testing and evaluation. This portal distributes all new geospatial data products (e.g., in-season crop cover, refined historical CDL) generated from this project. Like its award-winning precedent, CropScape, the iCrop web service system provides a suite of OGC web services and geoprocessing capabilities, including WMS and WCS data services, zonal statistics, PDF map production, change analysis, and on-demand data download.

Publications

  • Type: Journal Articles Status: Published Year Published: 2022 Citation: Lin, L., Di, L., Zhang, C., Guo, L., Di, Y., Li, H. and Yang, A., 2022. Validation and refinement of cropland data layer using a spatial-temporal decision tree algorithm. Scientific Data, 9(1), pp.1-9.
  • Type: Book Chapters Status: Published Year Published: 2021 Citation: Lin, L. and Zhang, C., 2021. Land Parcel Identification. Agro-geoinformatics: Theory and Practice, p.163.
  • Type: Book Chapters Status: Published Year Published: 2021 Citation: Zhang, C. and Lin, L., 2021. Image Processing Methods in Agricultural Observation Systems. Agro-geoinformatics: Theory and Practice, p.81.
  • Type: Conference Papers and Presentations Status: Published Year Published: 2021 Citation: Zhang, C., Yang, Z., Di, L., Lin, L., Hao, P. and Guo, L., 2021, July. Applying Machine Learning to Cropland Data Layer for Agro-Geoinformation Discovery. In 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS. IEEE.
  • Type: Conference Papers and Presentations Status: Published Year Published: 2021 Citation: Zhang, C., Di, L., Yang, Z., Lin, L., Zhao, H. and Eugene, G.Y., 2021, July. An Overview of Agriculture Cyberinformatics Tools to Support USDA NASS Decision Making. In 2021 9th International Conference on Agro-Geoinformatics. IEEE.
  • Type: Conference Papers and Presentations Status: Published Year Published: 2021 Citation: Lin, L., Di, L., Guo, L., Zhang, C., Di, Y. and Yang, A., 2021, Evaluating the Spatial and Temporal Accuracy of Crop Acreage Estimation from Different Agencies. In AGU Fall Meeting 2021. AGU.
  • Type: Conference Papers and Presentations Status: Published Year Published: 2021 Citation: Zhang, C., Di, L., Guo, L., Lin, L., Di, Y. and Yang, Z., 2021, December. Introducing Cyberinformatics Tool for Exploring Pre-and In-season US Crop Cover. In AGU Fall Meeting 2021. AGU.