This PDF file contains the front matter associated with SPIE Proceedings Volume XXXXX, including the Title Page, Copyright information, Table of Contents, and Committee Page.

Multispectral imagery captured from aerial robots, aircraft, and satellites provides data that may be leveraged to quantify crop nitrogen (N) status. While multispectral imagery is commercially accessible, there are few systems designed to transform multispectral imagery into actionable recommendations and prescriptions for farmers. N-Time Fertigation Management System (N-Time) was developed as an autonomous software decision support system (DSS) that analyzed multispectral imagery and applied a N recommendation algorithm to produce a binary action recommendation and accompanying prescription for an imminent fertilizer application via a center pivot irrigation system (fertigation). N-Time FMS automated the sensor-based fertigation management framework (SBFM) which utilized multispectral images of the crop canopy to inform the number and timing of fertigation applications made throughout the growing season. The framework involved a frequent (i.e. weekly) monitoring and recommendation cycle for which manual execution of computational processes for a single field required more than two hours. N-Time reduced the time requirement per field including user interaction to 7.4 minutes for 12 cm/pixel resolution multispectral imagery (i.e. UAV captured) and 3.1 min for 3 m/pixel multispectral imagery (i.e. satellite captured). N-Time also demonstrated greater than 99% agreement with computational outputs produced through manual execution with non-specialized software. During 2020 and 2021, N-Time was used to implement SBFM in research trials on approximately 320 ha of production-scale irrigated grain maize. These trials demonstrated that SBFM produced higher nitrogen use efficiency than the grower’s best management in 94% of implementations and higher profitability than the grower’s best management in 59% of implementations.

Volunteer cotton (VC) plants growing in fields of inter-seasonal crop like corn can act as host for the boll weevil pests; therefore, they need to be detected, located, and sprayed to prevent reinfestation of the pest in the following season. However, detecting the VC plants in corn fields has always been challenging as they remain hidden under the canopy and appear spectrally similar during the early growth phase. In this paper, we show that deep learning based YOLOv3 model can be used to detect VC plants in early growth corn field on RGB aerial images collected remotely by unmanned aircraft system (UAS) at a mean average precision (mAP) of 90.60% and F1-score of 86.35%. The approach of using deep learning to detect VC plants demonstrates its ability to be used for near real-time detection thereby expediting the management aspects of Texas Boll Weevil Eradication Program.

Cotton root rot (CRR) is a serious cotton disease primarily found in southwestern U.S., causing an average annual loss of about 29 million USD in Texas alone. Therefore, management of CRR infected cotton fields is crucial to the U.S. and Texas cotton industry. CRR usually appears at similar regions of the cotton fields each year, so detecting the locations of infected regions can make the management practices efficient for multiple growing seasons. Previous methods of mapping the regions of CRR involved classical image processing techniques like unsupervised machine learning methods, which are not viable for real-time detection. In this preliminary study, we present a deep-learning (DL) based method using YOLOv5 to detect the CRR infected regions of a cotton field, and then we demonstrate its ability for real-time detection by deploying it on an edge-computing platform (Pascal GPU of NVIDIA Jetson TX2 development board). In the end, we also show how the locations of detected CRR regions can be used to generate an optimal path for efficient management practices with the ant colony optimization (ACO) algorithm. Our preliminary results showed a moderate level of detection accuracy at a promising average inference speed of 11 frames per second (FPS). The total distance covered based on the optimal path of four detected regions of CRR was 160 m. Hence, through this study we were able to demonstrate that a DL based approach with the ACO algorithm has the potential to speed up management practices of CRR infected cotton fields with multispectral aerial imagery.

This paper presents the development of machine learning models and on-site infrastructure development for real-time plant health assessment in real-time using the images collected from unmanned aerial vehicles (UAVs). The data used for the training includes RGB images and spectral data collected from different UAVs equipped with RGB camera and multispectral sensors. Fully convolutional neural network (CNN), U-Net v2 was used to train the machine learning models. Statically defined geolocations extracted from the rectified rasters were augmented to efficiently generate a large dataset. The trained model was then used to interpret a greater variety of plant health information from just the RGB images. The paper discusses the development of a machine learning models to provide accurate, real-time, and actionable plant-level health information and eliminate the complexities involved with processing high density spectral data. The paper also discusses the development of stationary and mobile hardware required for real-time assessment of plant health.

In the last years smart irrigation has become an hot topic at the center of the scientific research attention providing to give solutions able to enhance these systems. The growing shortage of water resources, due to climate change and land use, requires more effective water use strategies for agriculture. Given the high consumption by the agricultural sector, we understand how making the consumption of water for irrigation more intelligent leads to a significant improvement in the management of this resource. In this paper we propose an IoT system composed of specific devices that cooperate in order to realize a smart irrigation management system able to correctly and efficiently use the water resource. The proposed architecture is based on sensor nodes flooded in the terrain and able to collect a series of data needed to operate the water irrigation throughout an intelligence decision maker based on a reinforcement learning algorithm modeled through a Markov Decision Process approach. The proposed reinforcement learning smart irrigation system has been compared with a system based on a simple threshold mechanism in terms of water and energy consumption.

Conventional stationary ground control points (GCPs) provide calibration data to unmanned aerial vehicles (UAVs) for georeferencing and calibration of images. However, conventional GCPs are inefficient in a large field setup. An autonomous mobile GCP robot (AMGCP) combined with a collaboration algorithm to communicate with the UAV was developed to improve the efficiency and accuracy of the UAV-based data collection process. The first AMGCP lacks robustness in its navigation and structural stability. This research focuses on improving the previous version of the AMGCP by using ROS to implement simultaneous localization and mapping (SLAM) navigation architecture and enhancing the structural stability with a suspension system and four-wheel drive power train. Preliminary tests of the improved AMGCP showed a robust navigation system and improved structural stability. The first AMGCP showed superior performance for georeferencing and radiometric, height, and temperature calibration than conventional stationary GCPs. The improved AMGCP added an intelligent navigation system and structural stability.

This paper presents an approach towards the autonomous collaboration between unmanned aerial vehicle (UAV) and an unmanned ground vehicle (UGV) for weed classification and detection. The paper outlines the development of an infrastructure for continued work towards a sustainable solution. RGB and multispectral data of a strawberry crop field infested with weeds were collected from UAVs. Fully convolutional neural network (U-Net v2) was employed in an effort to create a real-time capable onsite segmentation engine.

Using a combination of images collected from UAVs and highly accurate location information, target species were identified and isolated to efficiently generate a large dataset of images, which were then used to train a robust and state-of- the-art classifier. A highly sanitized dataset was used to effectively extract and augment a large amount of data without the need for manually labeling or outsourcing. The trained bounding box classifier were hosted onsite for real-time inferencing capabilities. In the proposed solution, a UAV surveys the areas of interest and transmits the images to the computing unit to detect and determine the presence of invasive weeds using the trained classifier. Positively identified locations by the UAV is stored for further investigation by the UGV. The UGV is equipped with high a precision IMU/GPS for the autonomous routing to the target location. After arriving at the target location, the UGV utilizes the onboard camera to confirm the presence of the invasive weed species. The UGV is equipped with a highly maneuverable robotic manipulator.

Unlike above-canopy imagery, under-canopy imagery is rarely used in agricultural research, but it can provide specific information about plants that above-canopy images may not carry, such as fruiting behavior as well as early nutrient deficiencies or diseases. In this research, under-canopy images of seven cotton varieties were collected with the goal of classification according to variety. An RGB FLIR camera was installed on a remotely controlled ground robot such that the camera was close to the ground looking upwards. A deep learning-based VGG16 network was used to extract features, and a softmax classifier in the network was trained to classify the images. The VGG16 network was able to classify the seven cotton varieties with 86% accuracy. The classification accuracy based on the area under the curve (AUC) of receiver operating characteristic (ROC) showed that the images belonging to the varieties CVTI-108, 110, 114, and 120 were the most accurately (AUC = 1.0) classified, while the images of variety CVT-115 were the least accurately (AUC = 0.93) classified.

Crop pest management has vast economic consequences vital to agricultural, natural lands, and even public health. In this work, we contribute a new low-cost UGV (unmanned ground vehicle) solution to solve this precision agriculture problem. This UGV is made of two electric bikes (bi-eBike) that are commercially off the shelves (COTS) and driven by a ROS (robotics operational system) compatible UGV. Our bi-eBike system offers a mobile platform that can be used as a mobile sensing service with multiple sensors such as multispectral cameras, microwave scanners, etc., as well as a mobile actuation/application service with actuators such as UV-C light insecticide, beneficial bugs, growth stimulant spreaders or sprayers, etc. By mapping and imaging plants in the field, farmers can treat individual plants instead of troubleshooting the entire field, reducing both their costs and negative environmental impact. This smart bi-eBike system can be supplemented with solar panels (photovoltaic) and a UAV (unmanned aircraft vehicle) landing/charging pad. Thus, one can expect that the SBB can have a long operational duration (10 hours or more) and large coverage of acreages where UAVs are used for variability mapping for site-specific treatment. This paper describes the system level concept, subsystem designs and integration, vehicle control electronics, autonomous navigation architecture, and some preliminary experimental results.

In remote sensing, the conversion of sensor recorded radiance for each pixel in a scene to surface reflectance is a first step in many diagnostic analysis tasks. This process, known as reflectance conversion, is vital in the production of accurate information for a variety of applications, one of which is precision agriculture. Several calibration methods have been explored in previous research and are widely used today, with two of these being the empirical line method (ELM) and the at-altitude radiance ratio (AARR). The AARR approach is attractive to remote sensing practitioners as it allows reflectance conversion to be carried out in real-time throughout data collection, accounting for changes in illumination conditions, which can substantially reduce collection setup and subsequent data analysis time/effort. Illumination changes during a collection greatly influence the recorded scene radiance, which can confuse subsequent analysis results. While ELM has been demonstrated to report lower error when compared to AARR, the error introduced is often times acceptable depending on the application requirements and natural variation in the reflectance of the targets of interest. An onboard, downwelling irradiance spectrometer integrated onto a small unmanned aircraft system as part of this research is utilized to characterize the expected error in generated reflectance at varying aircraft altitudes and is cross compared to the ELM approach. Although the error introduced by AARR is larger when compared to ELM, this study allows the reader to determine if the ease of collection afforded by this real-time reflectance conversion methodology produces sufficiently accurate data for their particular remote sensing application.

As the number of imaging systems continues to rapidly grow, so too does the need to perform spatial image quality control on captured data. While human assessment for general quality control may be utilized in some applications, target-based and scene-based automated spatial resolution methods may improve speed, accuracy, and efficiency. In certain applications, such as small unmanned-aircraft systems (sUAS) data collection, very large numbers of captured images may even make human assessment impractical, especially for multispectral and hyperspectral modalities. Imagery bands outside of the visible spectrum, such as infrared, may also prove more challenging for human assessment. Traditionally, scientific metrics utilized for spatial quality control such as MTF (modulation transfer function) were measured using in-situ targets such as the slanted-edge or Siemens star. However, newer advanced methods can now estimate these metrics without requiring dedicated targets. Automated machine vision methods utilizing techniques such as neural networks can provide these estimated measurement metrics for rapid image evaluation and quality control. MTF estimation without targets allows much more flexibility in assessment, enables a measurement map across each image, and eliminates the logistical burden of requiring in-scene fielded assets. We examine the performance of one such machine learning based MTF estimation system by placing checkerboard-type slanted-edge targets in multiple sUAS scenes including agricultural, urban, and forested subject matter. We then compare traditional target-based slanted-edge method measurements to the machine learning based estimated values and assess the overall accuracy.

This paper discusses the improved correlations between unmanned aerial vehicle (UAV)-based remote sensing and proximal sensor data. Better and increased correlation between remote sensing and proximal sensor data is necessary for the remote sensing data to be useful for precision agriculture. Portable ground control points (GCPs) were used for absolute positioning and will serve as an absolute georeference for the UAV data. RTK GNSS receiver and base station were used for centimeter level accuracy of ground truth data. This paper shows the results obtained from hyperspectral sensors that uses a high-performance GNSS/INS and multispectral sensor equipped with an RTK GNSS that is connected to an RTK GNSS mobile station for increased accuracy. The remote sensing data is used to calculate various vegetation indices including normalized difference vegetation index (NDVI), Green NDVI, and modified soil adjusted vegetation indices (SAVI). The indices are compared with the data obtained from proven proximal sensors that include Handheld Spectroradiometer and Chlorophyll Meter.

Vertical obstacles such as poles, towers, and wires present a significant hazard to low altitude flight in urban and tactical environments. The subtle protrusion of slender obstacles above surrounding terrain and buildings makes detection difficult. Current mitigation methods are not sufficient and necessitate an active sensor to enable obstacle avoidance. This paper surveys and assesses previous methods of vertical obstacle detection, including remote sensing, photogrammetry, and LIDAR sensors. Next, the paper examines the suitability of current point cloud segmentation methods, including clustering methods. Two new algorithms overcome unique aerial detection challenges. They broaden the safe volume by considering sparse data available from small obstacles at distances that will allow high flight speeds. Presented and existing clustering methodologies are evaluated against a variety of vertical obstacles in an Unreal Engine environment using Microsoft AirSim LIDAR simulation.

Estimating daily crop evapotranspiration (ET) is a critical component in tracking soil water availability for near real-time irrigation management. Energy and water balance models are two common approaches for estimating daily crop ET. These models informed with remotely sensed imagery can estimate crop ET spatially aiding both uniform and spatial irrigation management. A hybrid remote sensing-based ET model consisting of the two-source energy balance and water balance models was used in scheduling variable rate and uniform irrigation in maize and soybean fields. Variable rate irrigation was scheduled using two approaches: 1) a hybrid model informed with calibrated high resolution unmanned aerial system multispectral reflectance and thermal infrared imagery and 2) a water balance model informed with satellite multispectral reflectance imagery. The variable rate irrigation approaches were compared to uniform and non-irrigated approaches to quantify the effects on dry grain yield and net irrigation applied when managing variable rate irrigation using the hybrid and water balance models. A separate field study was completed to assess how well the remote sensing-based ET model could schedule uniform irrigations of an entire field.

Evapotranspiration (ET) is a crucial part of hydrological cycling, and its (ET) partitioning allows separate assessment of soil and plant water, energy, and carbon fluxes. ET partitioning plays an important role in agriculture since it is related to yield quality, irrigation efficiency, and plant growth. Satellite remote-sense-based methods provide an opportunity for ET partitioning at a subfield scale. However, one challenge is the resolution of the remote sensing data and relating these results to vine plants and irrigation subfield sections (valve units). With the small unmanned aerial system - sUAS such as AggieAir from Utah State University, ET and ET partition estimation via the two-source energy balance (TSEB) model at vineyards can be achieved. In this study, assessment of ET and ET partitioning, soil water evaporation and plant transpiration, using the Grape Remote sensing Atmospheric Profile and Evapotranspiration eXperiment (GRAPEX) information and aerial high-resolution imagery from AggieAir team will be performed. In specific, the performance of two implementations of the TSEB model: TSEB-PT (Priestley-Taylor approach) and TSEB-2T (dual temperature approach) on ET partitioning at a subfield scale will be evaluated, using ground and recently developed machine-learning-based Grapevine LAI imagery. The performance of the TSEB implementations will be documented for three major areas in California called Barelli, Sierra Loma, and Ripperdan, which encompasses major climatological regions in the state.

The application of modern technologies such as soil moisture sensors (SMS) provide an excellent water management and conservation tool, which reduces daily water use. The overall goal of this project is to employ UAV-based remote sensing technology to estimate bermudagrass performance and quality managed under SMS and ET based irrigation at Cal Poly Pomona’s Turfgrass Center. In addition to the UAV-based remote sensing technology, hand-held spectroradiometer and chlorophyll meter were also used to determine Bermudagrass performance. Data was collected from a UAV equipped with a multispectral sensor along with visual quality ratings and proximal sensor data. General linear model of correlation (p = 0.05) was used to analyze percentage green cover, density, NDVI, and other vegetation indices

Leaf water potential (Ψ_l) in vineyards and orchards is a well-known indicator of plant water status and stress and it is commonly used by growers to make immediate crop and water management decisions. However, Ψ_l measurement via the direct method presents challenges as it is labor and time intensive and represents leaf-level conditions for only a small sampling of the vineyard or orchard block. Models are existed for vegetation water status prediction by using optical and thermal images. Considering this, a small unmanned aerial system (sUAS) can potentially collect those data and help to build a predictive model at a high resolution. In this study, we identify relationships and trends of vineyard Ψ_l and sUAS imagery at different times of the day and throughout the growing season in California. This study examines aerial and ground measurements collected by the Grape Remote sensing Atmospheric Profile and Evapotranspiration eXperiment (GRAPEX) program over a period of eight years across California to build multivariable models for real time water status prediction. This preliminary analysis looks at spatial and temporal trends using a stepwise regression between leaf-level Ψ_l measurements and sUAS optical, thermal, and elevation data to identify potential predictors of Ψ_l, thus enabling mapping of Ψl at the scale of individual grapevines across the block. Such predictive models could be used to map the spatial variability in Ψ_l across multiple blocks during the growing season and at critical phenological stages in real time and improve the targeting of irrigation applications for vineyards and other perennial crops.

We have developed a field ready, fiber based snapshot imaging spectrometer for environmental monitoring and smart farming. The snapshot nature of the system enables monitoring of dynamic scenes without artifacts and portable, ruggedized opto-mechanics allows imaging from an airborne platform. We present hyperspectral imaging results of water stress observations in a laboratory experiment, field observation of sorghum under different growth conditions, and from an engineering flight.

Early estimation of canopy nitrogen (N) levels in cotton is necessary to maintain optimal canopy growth and derive best yields. Combining remote sensing and computer vision algorithms has made it possible to extract high-resolution spectral reflectance and canopy level morphology from larger fields. An experiment was conducted to study the response of cotton to four N application rates (0 kg/ha, 56 kg/ha, 112 kg/ha, 168 kg/ha). The differences in canopy N uptake and canopy height in response to the treatments became significant (p-value < 0.05) at the squaring stage. The objective of this study was to detect N stress levels in cotton canopy as early as the squaring stage and also estimate canopy N uptake (kg N/plant) from UAV-based multispectral images. Spectral vegetation indices and morphological features (canopy height and fractional canopy cover) were estimated from calibrated orthomosaics and digital elevation models (DEM). N uptake was estimated by single-parameter regression and multi-parameter regression (LASSO and Random Forest). Unsupervised k-Means clustering was used to separate the plots into three stress levels (stressed, moderately stressed, healthy) in a high dimension space. Canopy height was estimated with an overall R² = 0.911 and MAPE = 22.244, however the estimation errors were high before the squaring stage. Rededge (RE) band-based indices (NDRE and CI_rededge) were more sensitive to canopy N status than others. The best estimates of cotton plant N uptake from single parameter regression were obtained for measure canopy height (MAPE = 30.902) and NDRE (MAPE = 47.501). The multi-parameter models further improved N uptake estimation accuracy (LASSO MAPE = 22.133; random forest MAPE = 23.804). The results from k-means clustering showed that weekly changes (Δ_weekly) in features had better class separation and accuracy scores than instantaneous values of features observed at squaring. Clustering based on NDRE for canopy pixels and actual height gave the optimum separation (silhouette score = 0.488) and classification accuracy (adjusted rand index = 0.417).

Plant height is an important trait in crop breeding as it has high correlations with grain yield, biomass, and stress tolerance. Most of the studies so far for field-based high-throughput plant height phenotyping with UAS rely on ground control points (GCPs) for geometric calibrations due to the limited positioning accuracy of the GPS onboard. Setting up these GCPs is labor-intensive and is not feasible for a large field. The advent of commercial UAS equipped with Real-Time Kinematic (RTK) Global navigation satellite system (GNSS) technology are expected to achieve a centimeter-level positioning accuracy and have the potential of expediting and improving plant height phenotyping process. Hence, the objective of our study was to investigate the potential and dependency on GCPS of RTK GNSS enabled UAS technology and compare its performance to regular differential GNSS on plant height estimation. In the summer of 2021, images were corrected with two UAS – one with RTK GNSS; and the other with regular differential GNSS – for three different methods over cornfields: Method 1 used both the differential GNSS and the GCPs; Method 2 used both the RTK GNSS with the realtime corrections via cellular network and the GCPs; Method 3 used only the RTK GNSS with the real time corrections via cellular network but without GCPs. In this study, method 2 and 3 resulted with close accuracies, with a 𝑅² of 0.775 for method 2 and a 𝑅²of 0.760 for method 3 in maize plant height estimation. Method 1 had the lowest correlation (𝑅² = 0.250) in this study mostly due to the less data and data quality issues. Results from the study showed that the RTK GNSS enabled UAS has the potential and less dependency on GCPs in deriving accurate horizontal positioning and estimating plant height. Aerial surveys and plant phenotyping with RTK GNSS enabled UAS are more convenient and easily deployed in plant phenotyping and precision agriculture.

The demand for converting natural resources into human-consumed goods, mainly in terms of agricultural yield, is high due to both population and affluence growth. Agricultural yield per arable acreage is not optimal despite research advances in crop production (the so-called Green Revolution). Due to insufficient access to decision support systems, the technological advancements necessary to reduce the yield gap have had limited adoption. The purpose of this paper is to conduct a histogram analysis of the yield distribution per acreage to identify contributing factors of high yield and explain the differences in apparent efficacy of planting for different crops.

The scope of this research paper is limited to the cropland data provided by the United States Department of Agriculture's National Agricultural Statistics Service (USDA NASS). The USDA NASS provides county-level output per planted acre for several essential grain crops. In this paper, we will consider barley, corn, and alfalfa and evaluate the differences between these crops in terms of productivity from county to county and begin to assess the reasons for these differences. The analysis includes both quantitative (e.g., histogram entropy) and qualitative (e.g., case study) analysis.