Skip to main content
ARS Home » Plains Area » Sidney, Montana » Northern Plains Agricultural Research Laboratory » Pest Management Research » Research » Research Project #441824

Research Project: Geospatial Artificial Intelligence for Spatiotemporal Modeling of Pest Insects

Location: Pest Management Research

Project Number: 3032-22000-019-001-S
Project Type: Non-Assistance Cooperative Agreement

Start Date: Aug 31, 2022
End Date: Aug 30, 2024

Objective:
This project is intended to achieve three objectives: 1) To develop and deliver new spatially explicit deep learning-based methods and tools to address existing methodological gaps in geospatial-enabled agricultural pest research, 2) To extend deep learning-based geospatial methods into an uncertainty-aware statistical framework to better characterize and model spatiotemporal uncertainty in geospatial applications, 3) To apply the proposed developments to a long term record of grasshopper (Orthoptera: Acrididae) occurrence to forecast future insect pest outbreak potential in the Western U.S.

Approach:
We will develop new uncertainty-aware, GeoAI-based tools employing neural network algorithms, deep learning methods, and related machine learning tools to characterize and model complex spatiotemporal patterns and to demonstrate these approaches by assessing grasshopper outbreak risk across the Western U.S. We will collate multi-source grasshopper observation data representing grasshopper density documented over more than a 50 year period and then develop deep neural network-based GeoAI models and similar software tools to quantify associations of grasshoppers to a range of climatic, edaphic, and biotic variables. Deep learning methods employing neural networks, generative adversarial networks, and comparable machine learning algorithms often outperform more conventional techniques like random forest and generalized linear models when performing object detection and image segmentation. In addition to derived geospatial data characterizing environmental conditions (e.g., soil, drought, and precipitation grids), we will utilize multispectral imagery from sources like Sentinel-2 to quantify grasshopper and environmental correspondence at fine spatial and temporal resolutions using neural networks. Sentinel-2 optical satellite data is freely available for cloud-based computation and application development. Sentinel-2 data is available at 20 m spatial resolution and 1-2 week intervals, depending on seasonal cloud cover.