By Ce Yang, Xuechen Li, Alireza Sanaeifar, Aleksei Rozanov, and Bryan Runck, University of Minnesota
Golf courses experience significant winter damage due to harsh weather conditions, affecting turf health and overall playability. Traditional monitoring methods often fail to detect early-stage damage, leading to costly maintenance and recovery efforts. To address this challenge, AI-driven segmentation models and multispectral remote sensing could enable early detection of turf stress and provide insights into soil conditions, supporting proactive decision-making.
This research focuses on two key areas: 1) the use of multispectral imaging and AI for winter damage detection and 2) the application of remote sensing for soil moisture estimation. By combining spectral analysis, deep learning, and sensor-based soil monitoring, more efficient and data-driven maintenance strategies are being developed.
Winter damage detection with AI and multispectral imaging
Winter damage is difficult to detect due to its subtle and uneven distribution; for example, one section of a green might be completely unaffected, while a nearby location that doesn’t drain well could be completely killed from ice encasement. To improve accuracy, we are analyzing multispectral drone imagery using an AI-enhanced segmentation model. Vegetation indices such as NDVI, GNDVI, and NDRE (all of which are commonly used by turfgrass researchers) are integrated to highlight variations in turf health, making it easier to identify damage before it worsens.
We use an advanced AI segmentation model to enhance detection capabilities. This model can analyze large-scale turf data and understand relationships between different areas, making it more precise in identifying healthy and damaged turf. Additionally, it incorporates an attention mechanism (Squeeze-and-Excitation, SE) that dynamically highlights critical damaged areas, improving detection accuracy and efficiency. With these advancements, winter damage can be detected earlier so that action can be taken before issues worsen, leading to more effective turf maintenance strategies.
To enhance performance, synthetic data is generated via a Conditional Deep Convolutional GAN (cDCGAN), boosting data diversity and model generalization across courses and conditions. Field tests at the University of Minnesota in 2023 showed high accuracy. Compared to classic segmentation models such as U-Net, SegFormer, TransNet and DeepLabV3, our model demonstrates superior precision in differentiating healthy turf, moderate damage, and severe damage (Figure 1). These findings underscore the potential of AI-driven segmentation for proactive turf management.
Multispectral remote sensing for soil moisture estimation
In addition to turf health monitoring, multispectral remote sensing is being used to assess soil conditions. Spectral bands, including near-infrared, red, green, and red edge, are extracted to compute vegetation indices, which are then analyzed alongside soil sensor data such as soil temperature, volumetric water content (VWC), and dielectric constant. This integration provides a comprehensive view of soil health and moisture levels.
Strong correlations have been observed between vegetation indices—including GNDVI, NDRE, SAVI, SR, MSR, and MTVI2—and soil moisture content. These findings suggest that multispectral imaging can serve as a valuable tool for soil moisture estimation and predictive modeling.
Machine learning models have been applied to predict soil temperature based on these datasets and demonstrate the potential of combining remote sensing and AI to monitor soil conditions effectively, aiding in irrigation management and overall turf maintenance.
Spatial interpolation using Kriging approach
As an alternative to data-driven modeling, a range of methods exists to spatially interpolate ground sensor observations, one of the most popular being Kriging. In its essence, Kriging relies on variance change over distance to predict the missing values. By computing semivariance and creating empirical variograms, it’s possible to fit a theoretical model (Gaussian, Exponential, Spherical etc.) to it and then use this model for interpolation.
In the context of the current work, Kriging was used to interpolate 40 sensor readings (Figure 2) to a grid with ~50 cm2 spatial resolution across 8 separate timestamps (drone flights) to compare it with the aforementioned approach. This method can quite accurately predict soil temperature (Figure 3), but fails to adequately interpolate soil VWC and permittivity.
Research progress and future plans
Ongoing research is focused on refining multispectral image annotation to improve the accuracy of winter damage detection models. Enhanced training datasets and model calibration are expected to yield further improvements. Additionally, efforts are being made to optimize soil sensor deployment using UAV-based multispectral imaging for soil moisture estimation.
Future studies will involve the creation of pixel-based soil moisture maps using different sensor deployment scenarios. Maps will be generated using all available sensor data, interpolated data from half of the sensors, and data from only two diagonally positioned sensors. By comparing these approaches, the sensor deployment requirements will be evaluated, providing a scientific foundation for optimizing sensor density on golf greens.
Conclusions
The integration of AI-driven image analysis, multispectral remote sensing, and machine learning is revolutionizing golf course management. These technologies enable early detection of winter damage and more efficient soil monitoring, supporting proactive decision-making and targeted maintenance strategies. The combination of spectral imagery, predictive machine learning models, and optimized sensor deployment helps reduce recovery times, lower maintenance costs, and improve turf resilience.
As remote sensing and AI continue to advance, these methods will further enhance turf health monitoring and sustainability in golf course operations. Improved winter damage detection and soil moisture estimation will contribute to better resource management, ensuring optimal playability even under harsh winter conditions.