Skip to main content
ARS Home » Plains Area » Fargo, North Dakota » Edward T. Schafer Agricultural Research Center » Sugarbeet Research » Research » Publications at this Location » Publication #418075
Research Project: Improving Sugarbeet Productivity and Sustainability through Genetic, Genomic, Physiological, and Phytopathological Approaches

Location: Sugarbeet Research

Title: Detection of irrigation systems in North Dakota using the Mask R-CNN deep learning model and Landsat 8 imagery

Author
item BAZRAFKAN, ALIASGHAR - North Dakota State University
item PROULX, ROB - North Dakota State University
item Kim, James
item LIN, ZHULU - North Dakota State University

Submitted to: American Geophysical Union
Publication Type: Abstract Only
Publication Acceptance Date: 10/1/2024
Publication Date: N/A
Citation: N/A

Interpretive Summary:

Technical Abstract: The accurate detection of irrigation systems is crucial for effective waterresource management and agricultural planning. This study explores theapplication of the Mask R-CNN deep learning model to detect the locations ofirrigation systems across North Dakota using the panchromatic band ofLandsat 8, which offers a 15-meter resolution. A comprehensive datasetconsisting of 15 frames of Landsat 8 panchromatic imagery from the summerof 2024 was acquired and mosaicked using ArcGIS Pro 3.3.0. The datasetincluded 1,500 regions of interest, divided into training and test sets with an80% to 20% split, respectively. The Mask R-CNN model was trained undervarious tuning parameters, including different backbone models and epochs.Performance evaluation was conducted using training/test validation lossvalues. The results demonstrated that the integration of Mask R-CNN withLandsat 8 panchromatic bands achieved a detection accuracy of at least 75%for irrigation systems. This highlights the model's potential in large-scaleirrigation mapping using moderate-resolution satellite imagery. The results ofhyperparameter tuning show that the performance of ResNet150 in improvingthe accuracy of irrigation system detection surpassed that of other backbonemodels. Additionally, increasing the number of epochs from 200 to 300enhanced the prediction accuracy by 5%. Future work will focus on comparingthe effi cacy of alternative deep learning models and datasets, including thehigh-resolution National Agricultural Imagery Program (NAIP) imagery (60 cmresolution), using an expanded training and testing dataset. This comparativeanalysis aims to enhance detection accuracy and robustness, providingvaluable insights for precision agriculture and resource management.