We propose a solution with edge image processing and long-range connectivity named AICropCAM that can be used in drones, ground platforms, or as distributed sensor networks for plant phenotyping. We have successfully run multiple image classification, segmentation, and object detection models on this platform. Classification models help classify images based on image quality, crop type, and phenological stage. Object detection models could detect and count the number of plants, weeds, and insects and expand to count the flowers, fruits, and leaves. Segmentation models can separate the canopy from the background and potentially segment traits that indicate the nutrient deficit or disease. Canopy segmentation results help estimate leaf area index and chlorophyll content. Because the models run sequentially, like a decision tree, there is flexibility to select the most accurate model considering the crop type and the crop’s phenological stage that helps scan fields with multiple crops. The generated information is geo-tagged and transmitted through low throughput long-range communication protocol (e.g., LoRa) to cloud data storage. AICropCAM reduces 2-megabyte image files to around 100-byte actionable data, resulting in massive savings in data storage and transmission costs. This edge image capturing and processing system is open to improvement with new neural network predictive models and faster edge computers. This system provides plant scientists and crop breeders a low-cost, flexible phenotyping tool to extract multiple crop traits related to abiotic and biotic stress responses.

Maximizing crop yield while conserving resources is a pressing challenge in modern agriculture. Maize (Zea mays L. ), a staple crop worldwide, relies heavily on photosynthesis, making radiation interception a pivotal factor in crop growth and yield. This study presents a novel approach to improve maize crop productivity by harnessing the principles of phyllotaxy and optimizing planting patterns to efficiently intercept solar radiation. Through the enhanced 3D maize model generation algorithm, the simulation incorporates critical factors such as curved surface leaf area, leaf arrangement, plant spacing, and solar angles, allowing us to quantify the radiation intercepted by the maize canopy. We investigate the most efficient phyllotaxy for individual plants and evaluate a range of planting patterns and their impact on radiation interception at various growth stages. Simulation results reveal that optimizing planting patterns based on phyllotactic principles can substantially enhance radiation capture compared to traditional planting methods. In conclusion, our research signifies a significant step towards harnessing the power of plant arrangement and optimizing planting strategies, including the use of an enhanced 3D maize model, to maximize radiation interception in maize crops. The optimal phyllotaxis and planting patterns establish a distinct phenological target for breeders. These findings hold promise for the development of more resilient and productive agricultural systems in an era of growing global food demand and resource constraints.

As the world’s population grows and the demand for food rises, more attention has been paid to increase crop yields and enhance global food security. Modern remote sensing technologies enable us to capture spectral features (such as NDVI) of crop canopy, which are widely used to assess crop growth, health, and stress conditions. We noticed in the literature that crop NDVI shows short-term variation within a day. Therefore, in this study, we leveraged the Spidercam Field Phenotyping Facility at University of Nebraska-Lincoln to measure and quantify the diurnal variation of canopy NDVI for corn and soybean crops. The experiments and data collection were conducted in 2022 and 2023, with canopy reflectance measured by a spectrometer (400-1000 nm) at multiple days covering different growth stages. In each day, measurements were taken at multiple time points within a time window ±3 hours centered around solar noon. Our analysis showed a clear concave-shaped, diurnal trend in NDVI for both crops, with the lowest NDVI at solar noon. More analyses will be performed to quantify this diurnal pattern and dissect the sources of variation due to solar angle and change in canopy morphology. This research will further improve the accuracy and relevance of NDVI in plant phenotyping and many other scientific disciplines and applications.

Scoring plant phenotypes across large populations in multiple environments is a necessary precondition to both using natural genetic diversity to build genotype to phenotype models, study genotype by environment interactions and to carry out plant breeding to develop high yielding and more resilient cultivars. Here we explore data driven approaches using latent representations of leaf reflectance data collected from a large field experiment consisting of a subset of diverse maize lines drawn from the Wisconsin diversity panel (Mazaheri et al., 2019). In this experiment, 2 replicates of 752 inbred lines from the Wisconsin diversity panel were grown in field conditions. An ASD spectrometer was used to collect data on intensity of light reflected by leaves at 1 nanometer wide intervals between350 to 2,500 nm, resulting in a total of 2,151 reflectance intensity values measured for each plot. Two dimensional reduction approaches were evaluated for this dataset: conventional principal component analysis and an auto-encoder based neural network. Ten principal components were sufficient to summarize 99% of variance in the dataset. An autoencoder neural network comprising of an encoder having three dense layers and a decoder having four dense layers was able summarize variation within the dataset at a validation loss of 0.006 using 10 latent variables. A number of principal components and latent variables were correlated with several phenotypes quantified for a subset of the same field grown research plots (Figure 2A;2C). Chlorophyll, the major photosynthetic pigment in plant leaves, plays a substantial role in determining the overall pattern of reflectance for maize leaves. The abundance of chlorophyll was significantly correlated with PC2 (R2 = 0.31) (Figure 2B) which explained 11% of the total variance in higher spectral reflectance data. However, autoencoder based summary of the same trait dataset appears to have more accurately captured variation in chlorophyll abundance within this field trial with LV8 exhibiting a R2 = 0.59 (Figure 2D) with ground truth chlorophyll measurements. Both PCA and autoencoder based dimensional reduction captures a mix of variables which were heritable (i.e. a large proportion of total variance was attributable to differences between genotypes) and variables that were not heritable. Two of ten PCs evaluated exhibited H2 values >0.5 as did four of ten latent variables generated (Figure 3A; 3B). Genome wide association studies (GWAS) conducted using high heritability principal components and latent variables identified significant signals in 2 out of 6 cases (Figure 4A; 4B). Ongoing work is needed to evaluate the potential of using candidate genes underlying GWAS peaks to assign putative biological roles to latent variables estimated from raw sensor data by autoencoders or other dimensional reduction approaches.

Sorghum is an important cereal crop grown across the globe for its grain and biomass value. It can also efficiently use resources such as nitrogen, and multiple varieties that are nitrogen-use and light-capture efficient are constantly being developed. This study focuses on using the spectral signature of sorghum varieties to predict flowering days, which could be used as a proxy for plants’ growth/productivity and development trends, thus helping breeders make quick decisions about what varieties to move to the next stage. Multiple sorghum varieties from the sorghum association panel were planted in a replicate-design field experiment with the variable supply of nitrogen. The flowering days were monitored and recorded. The hyperspectral reflectance data were collected and used to build a sorghum flowering days predictive model. Although regression models such as partial least square have been used to predict plants’ phenotypes, the non-parametric ensemble machine learning model turned out to perform better on flowering days with an accurate model up to 5 days.