Grant Information
| Knowledge Area | Subject of Investigation | Field of Science | Percent |
|---|---|---|---|
| 404 - Instrumentation and Control Systems | 7410 - General technology | 2020 - Engineering | 80% |
| 402 - Engineering Systems and Equipment | 7410 - General technology | 2020 - Engineering | 20% |
Precision agriculture, or site-specific crop management, takes advantage of in-field spatial variability of many different cropping factors, such as soil type, soil pH, fertility, moisture content, crop vigor, disease, maturity status, and yield. In order to identify in-field spatial variability, many smart sensing systems are needed. Lee et al. (2010) reviewed different sensing technologies for specialty crop production, and reported that much improvement would be needed in various aspects of crop production. For example, "improved fruit detection algorithms are needed", and "more reliable, accurate, and rugged and less expensive sensing systems" will be needed.With the advances in electronics, more sensing technologies became available for crop production (Stone and Raun, 2015). However, still such sensing systems are not currently available for different cropping factors, and are very much in need by the growers. This project will focus on developing such sensing systems, so that the growers can easily adopt and utilize them for their crop management to increase yield and profit.
The following methods will be investigated and implemented for each of the specific objectives.To develop and test smart sensing systems for crop status including nutrient, water, disease, yield, and other crop information: Various sensing techniques will be used to develop smart sensing systems for crop production. Below are some examples of the proposed methods.Real-time multispectral imaging system using hyperspectral imaging: Many hyperspectral imaging applications have been conducted to identify important wavelengths to construct a multispectral imaging system for real-time field use. Using the results from the previous studies, multispectral imaging systems will be built for various applications. A commercial or custom-made multispectral camera will be used with different set of bandpass filters for different applications. The outcome of these systems will be maps of different crop status, such as diseased locations, water or nutrient stressed crop locations, yield information, etc.Micro hyperspectral imaging: Hyperspectral imaging is commonly used for many different agricultural applications; however, it is not true for micro or nano scale applications using hyperspectral imaging. A microscopic hyperspectral imaging system will be developed to investigate micro level phenomena, for example, crop disease progression, examination of differences in nutrient or water stress versus healthy crops. It is hypothesized that we can see and identify better what is happening at a micro or nano level.Thermal imaging: Thermal imaging will be utilized to develop various sensing systems along with different cameras such as a regular color camera, multispectral camera, depth camera, and hyperspectral camera. An imaging platform will be developed with these cameras for various applications. An autonomous vehicle will be used for unmanned imaging operations for various purposes.UAV application: UAVs will be utilized for many different applications. One example will be in-row weed control using machine vision and laser. A camera and laser unit will be mounted in an UAV for autonomous weed control. The UAV will be programmed to fly over the center of rows using pre-determined GPS coordinates. The camera will identify weed locations, and the laser unit will kill the weeds while flying close to crop rows. Another application will be to monitor crop diseases such as the citrus greening (Huanglongbing or HLB) or citrus black spot (CBS) diseases. A microprocessor and a multispectral camera will be mounted in an UAV along with a GPS receiver. Whenever diseased canopy or fruit are identified, the coordinates will be recorded for further management. The UAV will be programmed to autonomously fly to pre-determined locations in a grove.Laser weeding: An autonomous platform was built and tested in the Precision Agriculture Lab at the University of Florida for its performance. We will further identify proper laser power for controlling weeds in a crop field. The developed unit will be tested for transplanted crops.Real-time crop detection: We have worked for many years to develop immature green citrus fruit and blueberries, and developed various detection algorithms. Those will be combined and integrated into one platform for a real-time in-field crop detection system. Currently fusion of different cameras are investigated using a depth camera, thermal camera, hyperspectral camera, and color camera.Disease detection Citrus Black Spot (CBS) detectionA real-time in-field detection system for CBS will be developed and tested in a citrus grove. The developed technology will be transferred to the citrus growers in Florida as well as other citrus growing states in the U.S.Apple Marssonina Blotch (AMB) disease detectionCurrently hyperspectral images of AMB diseased apple canopies are analyzed to identify important wavelengths toward the development of a multispectral in-field disease detection system.There are many sensor technologies which are commonly used in other industries, but not in agricultural applications. Those include microelectromechanical systems (MEMS), field-programmable gate array (FPGA), wearable sensors, Internet of Things (IoT), and mobile technology. This project will actively search and adopt such technologies for precision agriculture applications.To combine the developed sensing technologies through multi-sensor data fusion and hardware integration toward the development of an integrated smart sensing system that can measure multiple crop information simultaneously: After individual sensing technologies are developed for different crop properties and status, the information from different sensors will be fused to produce better information and used to assist Florida crop growers in increasing yield and profit, making efficient management decisions, and reducing labor and input expenses. Sensor data will be fused across different sensors, time, and attributes. Various methods will be used for multi-sensor data fusion including neural networks, fuzzy logic, signal processing, statistical estimation, probability model, Bayesian statistical framework, maximum likelihood, linear Gaussian model, generalized Millman formula, Kalman filter, particle filter, ensemble learning, boosting and semantic alignment (Mitchell, 2007 & 2010).Multi-sensor integration will be also investigated to create a combined and integrated hardware system. For example, various cameras (such as multispectral, thermal, and regular color cameras) can be combined in a single platform to create an integrated imaging system to obtain more detailed information of an object of interest. A successful and effective image synchronization will be needed for this example. The integrated hardware system can be used for a single task or for multiple tasks simultaneously.The combined sensing system should provide better, more accurate and complete information than can be obtained by individual sensors. The integrated sensing systems should be cost effective, rugged and also be easy to use. Due to the current labor shortage situation, automated and mechanized sensing systems will be most appropriate for crop production to maintain the competitiveness and increase yield and profit.To disseminate the developed technologies to Florida growers through various means (commercialization, websites, extension publications, social media, and trade magazines) so that they could adopt the technologies, increase yield and profit, and maintain the competitiveness of crop production: One of the shortcomings from the previous CRIS project is the limited commercialization of the developed technologies. With the help of the Technology Licensing Office at UF, investors or sponsors for the technology will be actively sought through various channels such as vendor exhibitions at conferences, visit to the companies, phone calls, etc. Demonstration videos will be uploaded in the laboratory website through a Youtube channel for public access. As the University of Florida has an Extension publication (EDIS), the results will be published in EDIS so that growers will have access to the practical knowledge and technology. The results will also be published in trade magazines and related journals.
Target Audience
Strawberry and citrus growers
Changes / Problems
Nothing Reported
Training & Professional Development
Two undergraduate students were trained to learn more about precision agriculture, artificial intelligence, and sensing techniques for specialty crop production.
Dissemination Streams
The results of this project were disseminated through journal articles, conference presentations, and project reports.
Next Reporting Steps
A new Hatch project will be written to develop various sensing technologies for specialty crop production in Florida for the next five years.
Target Audience
Strawberry and citrus growers
Changes / Problems
Nothing Reported
Training & Professional Development
Nothing Reported
Dissemination Streams
The results of this project was disseminated by journal articles, conference presentations, and project reports.
Next Reporting Steps
We will continue the following projects. - Automated strawberry flower, immature and mature fruit detection and yield prediciton - Wetness detection of strawberry plants using thermal and color imaging - Two-spotted spider mite (TSSM) detection using color imaging and deep learning <br><br>
<br>What was accomplished under these goals? A region-based convolutional neural network (R-CNN) was used to accurately detect the number of strawberry flowers in images acquired using a ground-based imaging system. A modified VGG19 network was implemented and its performance was compared with other region-based methods such as the R-CNN, Fast R-CNN, and Faster R-CNN. Among these, the Faster R-CNN yielded the best detection accuracy of 86%. An UAV was used to acquire low-altitude (2 m and 3 m) images to detect strawberry flowers, immature and mature fruit from field images. The Faster R-CNN network was used for object detection, and yielded a mean average precision (mAP) of 0.83 for all detected objects. Distribution maps of flowers, immature and mature fruit were created which were the first of its kind. These maps can help growers harvest strawberries efficiently for increasing yield and profit. A machine learning method was developed to accurately detect the anthocyanin content in detached plant leaves (Arabidopsis) using color images. A quantile random forest model using sRGB color space performed the best with a coefficient of determination of 0.94 in predicting the actual accumulation of anthocyanin. <br><br><b>Publications</b><br>
Target Audience
Strawberry, blueberry and citrus growers
Changes / Problems
Nothing Reported
Training & Professional Development
Nothing Reported
Dissemination Streams
The results of this project was disseminated by journal articles, conference presentations, and project reports.
Next Reporting Steps
Nothing Reported
<br><br>
<br>What was accomplished under these goals? A ground-based imaging system was developed for detecting strawberry flowers toward yield estimation. Image processing algorithms were developed using color, shape, geometry, and R-CNN algorithm. The algorithms were tested in the field and yielded a 98% correct flower detection accuracy using a total of 100 images. Besides the ground-based system, images were acquired at different times over a growing season using an UAV to accomplish the same objective. Faster R-CNN method was used to identify number of strawberry flowers and yielded average precisions over 90%, comparing with manual flower counts. Yield prediction maps were created. Immature green citrus fruit detection algorithms were developed using multimodal imaging (color and thermal) and a new Color-Thermal Combined Probability (CTCP) algorithm. The CTCP value was used to confirm fruit regions in the color and thermal images. A total of 50 images were acquired from a citrus grove, and were evaluated for the algorithm. The precision of detecting fruit was 96% and the recall was 90%. For blueberry maturity detection, histogram oriented gradients (HOG) and color features were obtained from typical digital color images of blueberry bushes and were used to develop a Template Matching with Weighted Euclidean Distance (TMWE) classifier for distinguishing different blueberry maturity. The TMWE algorithm yielded correct detection accuracies of 95%, 92% and 84% for mature, intermediate and young fruit growth stages. <br><br><b>Publications</b><br>
Target Audience
Crop growers and general public who are interested in precision agriculture
Changes / Problems
Nothing Reported
Training & Professional Development
During the project period, one undergraduate student intern was hired to participate in various research activities. Also two graduate students were trained and facilitated with challenges, problem-solving opportunities, exercises on how to present and write reports and publications, and learning how to become an independent researcher.
Dissemination Streams
The results from this project have been disseminated to growers, researchers, and other stakeholdersthrough news articles, presentations, and journal articles.
Next Reporting Steps
Nothing Reported
<br><br>
<br>What was accomplished under these goals? A study was conducted to compare the performance of three different cameras (color, near-infrared, and depth cameras) in detecting immature green citrus fruit in a citrus grove. A circular object detection method, 'CHOICE', was proposed and compared with circular Hough transform. A deep learning algorithm (AlexNet) was employed to classify various objects in the images. It was found that the near-infrared images yielded the best detection results with a correct detection accuracy of 96%. A prototype yield mapping system for immature citrus fruit was designed, developed, and tested in a grove. It consisted of an autonomous navigation system and a fruit detection system with a color camera, an inertial measurement unit, and wheel encoders. A faster R-CNN network was trained and yielded an accuracy of 77% for fruit detection. A low-cost portable soil apparent electrical conductivity sensor was developed for use in mountainous areas and small farms. A single board computer (BeagleBone Black) was employed to collect georeferenced data at various user-defined signal frequencies and analyze them. The sensor was tested in a laboratory and a coffee field, and yielded good results. <br><br><b>Publications</b><br>
Target Audience
Citrus growers, blueberry growers, and general public who are interested in precision agriculture
Changes / Problems
Nothing Reported
Training & Professional Development
During the project period, two undergraduate student interns were hired to participate in various research activities. Also two graduate students were trained and facilitated with challenges, problem-solving opportunities, exercises on how to present and write reports and publications, and learning how to become an independent researcher.
Dissemination Streams
The results from this project have been disseminated to growers, researchers, and other stakeholdersthrough news articles, presentations, and journal articles.
Next Reporting Steps
Nothing Reported
<br><br>
<br>What was accomplished under these goals? IMPACT Precision agriculture, or site-specific crop management, takes advantage of in-field spatial variability of many different cropping factors, such as soil type, soil pH, fertility, moisture content, crop vigor, disease, maturity status, and yield. In order to identify in-field spatial variability, many smart sensing systems are needed. However, more reliable, accurate, rugged and less expensive sensing systems are not much available yet. This project focused on developing such sensing systems, so that the growers can easily adopt and utilize them for their crop management to increase yield and profit. Examples include immature green citrus fruit detection for early yield mapping far in advance of harvesting, and a novel detection system of citrus black spot disease using multispectral imaging. These are currently on-going projects and more results will be reported in later reports. ACCOMPLISHMENT During this reporting period (6/2/2016 - 9/30/2016), a study was conducted to develop a fusion method for color and thermal images to detect immature green citrus fruit on the tree. The method utilized photogrammetry and bundle adjustment to calibrate relative orientations of the cameras. The random sample consensus (RANSAC) method was used to find common points of interests from the color and thermal images, and a transformation matrix was developed to correctly register the images with an error of less than a few pixels. Another study was conducted to develop a machine vision system to evaluate the quality of harvested citrus fruit. Citrus greening disease (Huanglongbing or HLB), citrus rust mite and wind scar were included in the study. A graphical processing unit (GPU) and a deep learning technique was used to process videos of the fruit on a conveyer system, and yielded detection accuracies of 100%, 89.7%, 94.7%, and 88.9% for healthy, HLB, rust mite and wind scar, respectively. Citrus black spot (CBS) is another devastating disease in Florida, causing severely blemished and unmarketable fruit and yield loss eventually. An affordable vision based sensing method was developed to detect CBS infected citrus fruit under the field condition using two color cameras to acquire images in different channels (red, green, blue, and two near-infrared). These images were analyzed using morphological features to identify CBS lesions and acceptable accuracies were obtained. As an international research collaboration, detection algorithms were developed for the apple Marssonina blotch (AMB) disease from in-field hyperspectral images and a stochastic algorithm called particle swarm optimization (PSO). The disease causes early defoliation and eventually results in low quality and quantity of harvested apples. Ten important spectral features were identified to classify different degrees of infections. A support vector machine classifier was developed to distinguish healthy and diseased sample and 100% classification accuracies were obtained. Abundance estimation and spectral unmixing analyses were also conducted and reasonable separations were obtained among different classes (healthy, seemingly healthy and symptomatic). <br><br><b>Publications</b><br>