Recipient Organization
UNIV OF MINNESOTA
(N/A)
ST PAUL,MN 55108
Performing Department
Bioproducts & Biosystems Engineering
Non Technical Summary
Humans are faced with the great challenge of doubling agricultural production by the year 2050 to accommodate the food and biofuel needs of a human population expected to exceed 9 billion. Much of this increase in agricultural production will need to come from improved production efficiency, because of limited availability of additional arable land. Agricultural inputs (e.g., pesticides and fertilizers) are a necessary component of increasing agricultural production; however, the risk of these inputs causing environmental and human health impacts must be mitigated. To minimize the use of agricultural inputs while protecting yields, profitability and environment, remote sensing holds promise by quantifying spatial and temporal variability of the crop environment (soil nutrients, moisture, pest pressure, plant nutrient etc.) and guiding management decisions, such as targeted application of agriculture inputs only to areas of fields where needed. In addition, advanced remote sensing technologies help to address the challenges in efficient food production by high-throughput phenotyping. Remote sensing enables early season and real time detection of crop problems so that crop breeders can select the next best varieties for disease and stress resistance quickly. Remote sensing holds promise by quantifying spatial and temporal variability of the crop status and guiding management and breeding decisions. In animal production, remote sensing also gained popularity due to its capability of telling both physical and chemical features of products. For example, fertilized eggs are currently hatched without distinguishing female eggs from male eggs. Male chicks are culled right after they are hatched, which causes major economical and ethical concerns. Hyperspectral imaging has the potential of detection the gender of eggs before they are incubated.The overall objective of this project is about remote sensing in crop and animal agriculture. The focus is to explore new approaches utilizing unmanned aerial vehicles (UAVs) and advanced remote sensing techniques such as hyperspectral imaging and novel data analysis with advanced algorithms to solve agricultural problems including nitrogen management and recommendation for corn field and stress detection for wheat/barley disease resistance breeding, as well as in animal science, e.g. chicken egg gender detection before incubation. These are either emerging issues or future precision agriculture applications that farmers and researchers are eager to know more about. The reason advanced sensing technologies can help in tackling these problems is that with the development of advanced sensing technologies and increased computing power, remote sensing becomes more and more suitable in processing large amount of data efficiently for hundreds of acres of field at once, thus help growers make site-specific management decisions to reduce fertilizer, chemical and labor input, eliminate impacts on environment from agriculture, and increase profit. In addition, the advanced sensing techniques can be applied further into the area of animal science to help reduce operational cost and save resources such as eggs, which is of high nutrition value to human. The expected output of these remote sensing projects is a systematic solution for agricultural researchers and farmers to better use remote sensing techniques in solving fundamental agricultural problems. Toward this goal, we propose to study novel experimental and computational approaches integrating the following techniques: multispectral/hyperspectral imaging and machine learning. In addition to field experimental research, we plan to further develop and optimize the procedure of spatial and spectral data analysis so as to make it easy to adopt by both researchers and farmers.
Animal Health Component
50%
Research Effort Categories
Basic
20%
Applied
50%
Developmental
30%
Goals / Objectives
The overall goal of this project is to explore the use of remote sensing in agriculture and high-throughput phenotyping. The focus is to experiment new approaches utilizing unmanned aerial vehicles (UAVs) and ground vehicles, as well as advanced remote sensing techniques such as hyperspectral imaging and color imaging to solve agricultural problems including corn nitrogen deficiency detection and management, wheat/barley disease detection for high-throughput breeding. These are what growers and researchers are eager to know more about. The reason remote sensing can help in tackling these problems is that with the increased computing power, remote sensing techniques become more and more suitable in processing large amount of data efficiently for many crop varieties in large fields at once, thus help the world in agricultural production. The expected output of these remote sensing projects is a systematic solution for agricultural researchers and growers to better use remote sensing techniques in solving fundamental agricultural problems. Toward this goal, we propose to study novel experimental and computational approaches integrating the following techniques: multispectral/hyperspectral imaging, machine learning and deep learning. In addition to these research interests, we plan to further develop and optimize the remote sensing protocol for nitrogen management and disease phenotyping by multi-year study so as to make it easy to adopt by both researchers and growers. Considering the large potential of using remote sensing in animal agriculture and need of preventing more chicken culling, we will also test the capability of hyperspectral imaging for in-vivo egg sexing to save half of the fertilized eggs from being hatched and male chicks being culled. Upon this overall objective, we have the following three objectives to be achieved.
Project Methods
1. Corn nitrogen management using hyperspectral imaging and crop modeling.All field research will be conducted at the Ecology and Agricultural Research Farm (AERF) of the University of Minnesota in Waseca County, MN. The AERF site is approximately 24 ha and consists of nine small rain-fed fields (0.8-2.4 ha in size each), with clay loam soils (Webster and Nicollet) and slopes of 1 to 3%. Each field at this site is equipped with independent subsurface drainage systems that were designed to have two drainage intensities; conventional intensity plots were designed to have drainage rates of 1.3 cm/day, while high intensity plots were designed for a drainage rate of 5.1 cm/day. One of these fields (subfield 2) will be used to set up a small N rate trial that will be used to develop spectral response to N and yield response to N curves to be used for determining variable-rate sidedress applications. This subfield will also be used for calibration and validation of the EPIC model. The treatments of subfield 2 will be set up with two replications as a randomized block design, using N fertilizer treatments. The area of each N rate treatment plot will be approximately 14 m x 30 m, as the plot size must be sufficiently large to accurately record yield with a combine yield monitor.?Multispectral and hyperspectral imagery can enhance the detection of leaves with N deficiencies. On the data analysis side, it is noted that there are more than 200 partially redundant bands in a hyperspectral image, which require post-processing to improve prediction ability. We will conduct various algorithms to achieve majorly enhanced corn N prediction results. Furthermore, unlike the traditional approach of only using spectral information, we will also utilize the spatial information from the collected images.The EPIC model relies on detailed field-scale inputs, such as weather data, topography, soil attributes, and agronomic field management practices. Soil attributes such as soil moisture, soil organic matter content, and soil texture exhibit considerable spatial variability. Obtaining accurate information about this variability is time consuming and costly. The USDA NRCS provides online access to the Soil Survey Geographic (SSURGO) database (https://www.nrcs.usda.gov/wps/portal/nrcs/detail/soils/survey/?cid=nrcs142p2_053627), which contains information about spatial variability in soil organic matter and soil texture at the scale of 1-5 ha units nationwide. EPIC is designed to access this database, and then estimate spatial patterns in soil moisture and crop biomass (Leaf area index) based on daily precipitation, temperature and other climatic data.2. Field wheat FHB disease detection using color imaging and deep learning algorithmsA diverse array of genotypes known to differ in their reaction to the disease will be planted in the field at St. Paul and Crookston, MN in both 2020 and 2021 to evaluate the feasibility of image processing for assessing the severity of FHB in small grains. Two-row plot design will be adopted for easy phenocart access to the field. The experiments will be planted in a randomized complete block design with six replications. Inoculations will be made using the "grain spawn" method, where equal amounts of two to six regional F. graminearum isolates (provided by Ruth Dill-Macky) will be applied uniformly throughout the nurseries to initiate FHB. The first inoculation will be made when the flag leaves of the earliest maturing plants are expanding. One to three successive inoculations will be made at about one-week intervals to ensure that sufficient inoculum is available for infection of later maturing accessions. To create sufficient moisture on the spikes for optimal FHB infection, an overhead irrigation system will be used. Irrigation will begin after the first inoculation and continue until the latest maturing accessions reach the late dough stage of development. Using these methods, we have achieved successful FHB nurseries during each of the past 20 years.For color images, mask RCNN model can be adopted for spike detection. A test dataset (450 samples) has been processed and a total of 2127 spikes were detected and extracted. The full dataset with more than 1000 images will be processed next. After that, the gray differences between diseased and healthy pixels can be used for the detection of the FHB areas of each spike using conventional segmentation algorithms of machine learning. The Modified Region Growing algorithm will be used to segment the diseased areas from the entire spike. The K-means and OTSU algorithms will also be implemented to process the gray spike images used for the comparative tests. Eventually, small areas will be treated as noise and removed. Hyperspectral images will also be captured from the platform by a Resonon Pika II hyperspectral camera on the same days.3. In-vivo chicken egg sexing by lab hyperspectral imaging.The proposed study will start with scanning white eggs before incubation. Experiments with 200 eggs will be carried out following the image collection protocol between Day 0 and Day 4 of incubation. The line-scanning hyperspectral camera carried by a track will move from one end of the egg tray to the other end to obtain a spectral image cube (two spatial dimensions and one spectral dimension) that contains both spectral and spatial information of the eggs. Eggs will be placed in an egg tray that provides the best transmittance imaging. The egg tray will allow enough backlight to be used on at least 1/6 of the eggshell surface for hyperspectral transmittance imaging. A black rubber ring will be placed on the edge of the individual egg holder in the tray to prevent light leaking during transmittance image collection. A semi-transparent white standard will be placed on the side of the egg tray for image calibration. After the first round of scanning before incubation, the second round of scanning will be at the end of Day 1 of incubation and the third round on Day 2. As the incubation goes on, more and clearer blood vessels will develop, providing spectral information that was proved to be useful in detecting sex in previous studies. Therefore, there will be a fourth and fifth round of scanning on Day 3 and 4 of incubation accordingly.Hyperspectral images provide both spatial features and spectral features for image analysis. Image pre-processing procedures such as reflectance calibration, spatial and spectral smoothing and edge detection will be carried out on the images. These procedures improve the image data by suppressing noise and enhancing color and morphological features. Hyperspectral imaging is a sophisticated technique that provides more useful information than many other sensing techniques. However, hyperspectral cameras are costly and the image processing with hundreds of bands requires significant computing capacity and time. Therefore, spectral band selection needs to be done to insure high speed computation, storage space-saving, simplicity in further statistical analysis and easy adoption by the industry. Algorithms such as ensemble feature selection, sparse support vector machines and Bayesian methods will be used to analyze the hyperspectral images with the aim of obtaining the most useful wavebands (features) for classifying female and male eggs for various incubation dates. It is expected that 910nm ranks the highest in the selected wavebands, although eggs on different days of incubation may need slightly different sets of wavebands for the classification as the embryos develop. The Receiver Operating Characteristic (ROC) curve will be analyzed to optimize the number of wavebands and receive the best classification results. The earlier in the hatching process sex determination is made, the more likely male eggs can be repurposed, which will improve economic returns. Software to be used for image processing and decision making includes Matlab 2018b and ENVI.