This session will feature academic-based presentations on image classification approaches including AI and machine learning.
11:00 – 11:15 AM – Advancing Growth and Yield Predictions for Pinus taeda (L.) Plantations: Integrating Lidar-Derived Metrics and Machine Learning Models Across Diverse Sites
This research examines the accuracy of two machine learning (ML) models, Random Forest (RF) and Support Vector Machine (SVM), in predicting tree stem volume over four years for 8-year-old Pinus taeda (L.) plantations using lidar-derived metrics from 2017. The study incorporates Individual Tree Crown (ITC) metrics and Competition Indices (CI) into these models. Data from two locations in Virginia and North Carolina, with varying silvicultural regimes and planting densities, were used. Field measurements from 2021 served as the validation dataset. The findings show that both RF and SVM models can accurately estimate stem volume with normalized root mean square error (nRMSE) values below 15%, confirming the hypothesis of acceptable accuracy. The RF model accounted for 49.16% of the variance in the data with an R² of 0.47 and nRMSE of 10.86%, while the SVM model achieved an R² of 0.59 and an nRMSE of 9.59%. The importance of various lidar-derived predictors was evaluated, and models were streamlined to the top 7 predictors, preserving predictive performance while simplifying the models. The top predictors included variables from both ITC metrics and CI. The simplified RF model achieved an R2 of 0.37 and an nRMSE of 11.88%, and the simplified SVM model achieved an R2 of 0.55 with an nRMSE of 9.14%. Furthermore, the streamlined models were tested on different plantations in Texas with trees aged 7-9 years. Sample plot measurements from 8 stands were collected to estimate stand-level tree stem volume, serving as an additional validation set. Variations in model accuracy across different study sites and planting densities were noted, underscoring the impact of local environmental conditions and stand characteristics. This study highlights the potential of integrating lidar-derived metrics and ML techniques in forest growth and yield models, providing valuable insights for forest management practices. Future research should consider additional environmental variables and advanced ML techniques to enhance model robustness and applicability.
Gunjan Barua, Virginia Tech
11:15 – 11:30 AM – Classifying Hydrilla (Hydrilla verticillata) Invasions in Florida Lakes from Sentinel-2 Imagery using a Custom Vegetation Index with Machine Learning Techniques
Submerged Aquatic Vegetation (SAV) are vascular macrophytes rooted in bottom sediments with leaf shoots completely submersed under water. SAV are primary producers on a trophic scale providing habitat for other microbial and faunal communities and are important drivers in biogeochemical cycling. Hydrilla is an invasive SAV, which can rapidly expand into large monospecific infestations competitively excluding other native macrophytes and disrupting whole lake ecosystems. In Florida, hundreds of acres can quickly expand to tens of thousands of acres within a couple of years. Â As a proactive measure, millions of dollars are spent annually for managing hydrilla at various stages. Effective hydrilla management is contingent on accurate and timely monitoring typically performed by field surveys which can be time consuming and low resolution. This study utilizes Sentinel-2 (S2) satellite imagery for developing a remote sensing approach to broad-scale monitoring of hydrilla. Random forest models were trained with all 13 spectral bands of the S2 electro-optical sensor to classify hydrilla from 2021 imagery and complementing ground-truth data collected on Lakes Yale and Apopka in Florida, USA, with over 3,000 and 10,000 acres of hydrilla infestation, respectively. We developed a customized three-band index, the Submerged Vegetation Index for Hydrilla (SVIH), which can effectively distinguish live submerged hydrilla from open water, planktonic algae and other emergent macrophytes. Ground-truth validation of SVIH produced a high F1 score distinguishing hydrilla, but with a low recall suggesting a limit to the detection of sub-surface hydrilla masked by water. These findings show progress towards the development of a comprehensive, high frequency surveillance platform for early detection of incipient invasions, monitoring growth expansion of infestations and evaluating performance of treatment suppressions.
Ayesha Malligai M., University of Florida
11:30 – 11:45 AM – Mapping Residual Dry Matter across California Rangelands with UAV LiDAR and Hyperspectral Imagery
Residual Dry Matter (RDM) is the non-green/non-photosynthetic plant material left on the ground at the beginning of the growing season in rangelands. It is a landscape-scale estimate of aboveground biomass (lbs/acre or kgs/ha) used by agencies to guide grazing and fire fuel management across the Western United States. Estimating RDM through traditional field methods is labor, cost, and time-intensive, making large-scale sampling challenging. Studies have suggested using remote sensing to quantify non-photosynthetic vegetation biomass across rangelands and grasslands, however, little is known about the accuracy and applicability of  remote sensing technologies to quantify RDM directly. Portable hyperspectral sensors capture biophysical properties such as cellulose and lignin content of dry matter with hundreds of narrow spectral channels, while UAV LiDAR provides centimeter-scale 3-dimensional structural data. This research involves correlating RDM ground-reference measurements with co-located UAV LiDAR and field spectra across the Jack and Laura Dangermond Preserve, a 24,000-acre preserve with extensive grazed and ungrazed rangelands on the Central California Coast managed by The Nature Conservancy. Random Forest Regression models indicate strong relationships between RDM and UAV LiDAR data (R2 = 0.90; RMSE = 8.49) and Multiple Linear Regression models indicate a moderate relationship (R2 = 0.50; RMSE = 18.61). Random Forest Regression models indicate strong non-linear relationships between RDM and hyperspectral data (R2 = 0.87; RMSE = 8.95), but weak linear relationships (R2 = 0.22; RMSE = 21.77) when evaluated with Multiple Linear Regression Models. Combining hyperspectral and LiDAR data results in slightly decreased Random Forest (R2 = 0.88; RMSE = 8.60) and Multiple Linear Regression (R2 = 0.50; RMSE = 17.47) model performance than using LiDAR data alone. The potential outcomes of this research include improved fine-scale fire fuel mapping and grazing monitoring, as well as exploring the value of field-based hyperspectral sensors and UAV LiDAR for real-time RDM quantification and forage monitoring.Â
Authors of this presentation include Bruce Markman (San Diego State University), Dan Sousa (San Diego State University), Lloyd (Pete) Coulter (San Diego State University), Janet Franklin (San Diego State University), Scott Butterfield (The Nature Conservancy), Moses Katkowski (The Nature Conservancy)
Bruce Markman, San Diego State University