Location: Range Management Research
Project Number: 3050-21600-001-010-A
Project Type: Cooperative Agreement
Start Date: Jul 1, 2024
End Date: Jun 30, 2026
Objective:
Satellite-derived maps of rangelands are now a bedrock of rangeland management and research. Technological advancements in the last decade have reduced previous computational limitations, allowing scientists to produce maps of various ecosystem indicators, metrics, and processes. These maps are used in numerous operational activities, including rangeland monitoring, conservation planning, and management/treatment evaluation. At a national level, the forward looking strategies and visions produced by many agencies frequently rely on maps as the foundation. Furthermore, research activities have also been enhanced by satellite-derived maps, often unlocking new analyses and findings by providing data across broad regions and time periods.
Continued investment into satellite remote sensing has produced new datastreams that can improve estimates of rangeland metrics. Synthetic aperture radar (SAR), spaceborne lidar, very high resolution satellite imagery, and very high resolution aerial imagery are now widely available, complementing the medium resolution optical imagery traditionally used. Recent advancements in machine learning and artificial intelligence provide various approaches to utilize these datastreams to the fullest extent, discovering and leveraging relationships that were largely unfeasible or impractical just a few years ago. Moreover, remote sensing foundational models are becoming a working reality, making it easier than ever to leverage remotely sensed data and obtain high quality results with just a small number of samples, i.e., few-shot learning.
The primary objective of this agreement is to develop, investigate, and test methodologies to improve satellite based monitoring of North American rangelands. The primary focus will be to improve estimates of fractional cover; secondary will be to improve remote sensing foundational models for rangeland applications. The cooperator will perform various experiments to test methodologies. Results will be published in preprints and peer-reviewed journals. Materials used in the development and testing of approaches will be made available to the ARS PI.
Approach:
To improve estimates of fractional cover, various new methods and strategies will be explored. Previously successful, convolutional neural networks (CNNs) remain a practical and usable option; exploration of new architectures and frameworks, along with additional data inputs and updated training data, will likely reduce error and improve estimates. More recent approaches, including vision transformers or video vision transformers (ViTs, ViViTs), are better suited for modeling complex interactions, and also better learn relationships between more distant data. Furthermore, when combined with a self supervised learning, transformers can easily learn representations that include multiple sources of data, as well as handle missing data. Learned representations can then be used downstream in numerous applications.
The cooperator will use combinations of the various approaches discussed above to improve satellite based monitoring of rangelands. The cooperator will begin with less complex methodologies (i.e., CNNs) and work their way to more complex approaches (i.e., ViTs, masked autoencoders, etc.). The cooperator will communicate frequently and consistently with the ARS PI to ensure that objectives are being met.
