Earlier research & posters

The ever-growing increase in population and uncertainty in approaches administered by farmers mandate more systematic ways to implement agricultural management practices. Precision agriculture can address this need by providing site-specific solutions that maximize profit and minimize cost. This research is aimed at evaluating snap bean yield, prior to harvest, using UAS-based hyperspectral imagery, and identifying discriminating wavelengths that best explain snap bean yield.

This study compares UAS-based structure-from-motion (SfM; photogrammetry) and LiDAR point clouds collected concurrently across five snap bean field flights, evaluating the suitability of SfM for crop phenotyping. Cloud-to-mesh distance maps and multiscale model-to-model cloud-comparison maps quantify differences between point clouds, digital elevation models, and crop-height models. Crop height and row width derived from crop-height models are also compared with field measurements. Both SfM and LiDAR achieve an average RMSE of approximately 0.02 m for crop height and 0.05 m for row width. The analysis shows that SfM can be comparable to LiDAR under the same UAS flight settings, although its altimetric accuracy depends strongly on the number and distribution of ground-control points.

This field study extends earlier greenhouse work to low-altitude UAS hyperspectral imagery for snap bean yield prediction. Two years of observations from 2019 and 2020, covering 24 and 18 plots across two upstate New York locations, support early- and late-harvest models for pod weight and seed length. The workflow uses reflectance calibration, vegetation extraction with RENDVI and Spectral Angle Mapper, spectral denoising, partial least squares regression, and Jostar feature selection. Genetic Algorithm, Ant Colony Optimization, Simulated Annealing, and Particle Swarm Optimization rankings identify informative wavelengths, while ACO and SA perform best among the tested methods. Pod-weight models achieve R² values of 0.78–0.93, and seed-length models achieve R² values of 0.83–0.98. The analysis identifies 44–55 days after planting as an effective period for yield assessment and supports future transfer to operational multispectral systems.

This greenhouse study evaluates machine-learning methods for snap bean growth-stage and pod-maturity classification. Four growth stages and six pod-maturity levels are assessed using VNIR–SWIR (400–2500 nm) reflectance collected with an in situ spectroradiometer. Binary growth-stage classification reaches 90–98% accuracy, with continuum removal providing the best enhancement method. Multi-class growth-stage classification identifies 493 nm and 640 nm as informative wavelengths and reaches approximately 79% accuracy. Continuum-removed spectra distinguish ready-to-harvest from not-ready-to-harvest pods at approximately 78% accuracy. The study provides a foundation for testing a compact UAS-mounted multispectral sensor with fewer than ten spectral bands in operational snap bean production.

This controlled greenhouse study investigates relationships between snap bean yield and plant spectral and biophysical information collected with a 400–2500 nm hyperspectral spectroradiometer. Observations from 48 individual plants across a 69-day growth period support partial least squares regression and cross-validation of early- and late-harvest yield models. The selected spectral bands are intended for transfer from a hyperspectral system to a more cost-effective multispectral sensor. The best model achieves an RMSE of 3.02 g per plant and an R² of 0.72. Optimal spectral-evaluation periods occur 20–25 days before early harvest and 32 days before late harvest, supporting near-real-time yield estimates from a UAS-mounted wavelength-programmable multispectral imager.