Improved crop row detection with deep neural network for early-season maize stand count in UAV imagery
Introduction
Stand counts in crop production refer to counting the number of plants grown in a specific area, in other words, obtaining the population of plants. In-season crop stands are often different than the planting population. Early season crop stands are usually a result of seed quality, planter performance, and seedlings’ response to soil and weather during emergence; while mid to late season crop stands are usually a result of weather, soil, fertilizer, pest pressure and other management practices. Stand counts are one of the most common ways farmers assess plant growth conditions and management practices throughout the season and can be used to make management decisions, such as determining if replanting is required. It is also important for crop insurance industry, planter manufacturers and other agricultural professionals and companies to evaluate products and management practices.
The traditional way of conducting stand counts is to walk in the field, manually sample a few areas and estimate the total number of plants in the field based on the results in sampled areas. This method is time-consuming, laborious, and does not cover the entire field. Efforts have been made over decades to develop automatic ways to replace the labor-intensive manual stand counts. Satellite or manned aircraft based remote sensing systems had been used to cover large area. However, due to the low spatial resolution (3–5 meters) obtained from those remote sensing platforms, plants were not clearly visible in resulting images. Because of this, only spectral reflectance, rather than morphological or texture features, were used to capture the signal of non-uniform stand distributions (Erickson et al., 2004). Ground-based systems had been developed for early and mid-growth stage stand counts when plants are small. Most of them used the machine vision technology to either look downward or from the side between rows (Luck et al., 2008, Thorp et al., 2007, Nakarmi and Tang, 2012, Shi et al., 2013). Although they achieved good counting accuracies with very high-resolution data, currently the throughput is still not high enough to practically apply such technology to large fields.
Stand counts for row crops such as maize are most valuable early in the season when farmers still have time to replant the field if the stand is poor enough. Knowing about a poor stand later in the growing season is not as helpful, as it is too late for action to be taken. Therefore, a system that operated on V4-5 growth stage maize would have a distinct advantage. With the advances in unmanned aerial vehicle (UAV) technology, it becomes possible to have relatively high-resolution data (spatial resolution of 2 to 25 mm) and cover a relatively large area (5 to 25 hectares) almost whenever needed in agriculture (Shi et al., 2016). Though the spatial coverage is still limited by flight endurance, UAV technology nowadays is feasible to provide coverage for an entire field. How to make use of the highly detailed images collected by the UAVs to help agricultural production is the research question to answer. Fortunately, the advancements in the machine/deep learning area provide many possible and powerful solutions. Su et al. (2018) used various vegetation indices derived from a UAV-based five-spectral-band multispectral camera and Random Forest with Bayesian hyperparameter optimization to detect wheat yellow rust. Lin et al. (2017) proposed that region-based convolutional neural networks could improve the accuracy of strawberry flower detection to reliably predict fruit yield. Zou et al. (2019) reported a method using UAV, hyperspectral image sensors, and machine learning based data processing algorithms to accurately measure tree height and diameter at chest height in forests to assess the growth rate of cultivars. Etienne and Saraswat (2019) used unmanned aerial vehicles and image sensors to capture image data of maize and soybean fields to automatically detect weeds in the fields at different stages of the growing season using machine learning techniques. Developing an economical solution for monitoring plant health has always received attention from different aspects. Do et al. (2018) developed a method for monitoring nitrogen and water in plants by analyzing digital images of citrus plants collected from UAV. They evaluated several machine learning techniques such as simple linear regression and convolutional neural networks, and the results were all satisfactory.
For in-season crop stand count specifically, applications that used UAV and machine learning techniques had been proposed as well. Kitano et al. (2019) proposed using deep learning on images captured from UAV for maize stand count. After splitting maize and ground by recognizing each plant in the image using U-net (Ronneberger et al., 2015), they used an opening morphological operator to separate the plants that were close together. Finally, the blob detection (Danker and Rosenfeld, 1981) method was applied for quantifying the number of maize plants. However, their method relied too much on accurately detecting individual plant. If the image resolution is reduced (usually due to the increased flying altitude of the UAV), for example, the width of the plant is reduced from 20 pixels to 5 pixels in the image, this approach would fail due to the small size and ambiguous features of the individual plant. Varela et al. (2018) also proposed a method for maize stand count. Unlike Lin’s approach, they first used the Excess Green (ExG) index (Yaakob, 2017) to separate the vegetation from the background using the difference in color. Then they used the Canny Edge detector and Hough Transform method to detect the rows of plants and the orientation angle, which avoided the interference caused by weeds (they assumed that the weeds were located at inter-row areas). Then, they used 10 geometric descriptors to describe the characteristics of the plant and used a decision tree classifier to identify individual plants and count the number of them in each row. However, similar to Lin’s work, Varela’s scheme relied on detecting each individual plant as well, therefore the performance would also be reduced as the image resolution decreases. In addition, since the directions of crop-rows in the images usually are not absolutely horizontal, the authors first calculated the rotation angle required to rotate to the horizontal level for each row, and then voted to select an appropriate value for a whole image. This method was based on the assumption that all the plant rows in one image were almost parallel. In real cases, rows may not be perfect straight and parallel due to terrain and planting.
Therefore, in this study, a novel early-season maize stand count approach is proposed, which has the following advantages: (1) To better conform to various terrain, each crop-row is considered as an instance, and the stand count algorithm is customized for each row. An innovative instance segmentation algorithm – MaxArea Mask Scoring RCNN is developed for detecting the crop-rows, regardless of their directions and shapes displayed in the image including curved rows that intersect at different angles with numerous gaps. (2) Instead of detecting individual maize plant in a row, a simple but efficient sparse region detection and stand count estimation strategy is proposed, which utilizes the property of connected components in image domain. (3) It is a robust system, which automatically adjusts the parameters of the model according to UAV’s flying altitudes and performs stably. (4) The processing speed of this system is fast, which provides the potential for real-time applications in the future.
Section snippets
Materials and methods
Fig. 1 shows the main steps of the proposed system. First, the field map of an entire maize field is stitched from many small raw images, which are collected from a relatively high altitude using a UAV equipped with a high-resolution camera. Second, since the size of this map is large, it exceeds the ability of a general computer processing data using neural networks, therefore it is necessary to divide the map into several patch images. The number of patch images can be calculated based on the
Evaluation of crop-row detection
Average Precision (AP) (Yilmaz et al., 2016) is usually used to evaluate the performance of object detectors. It refers to the average of the highest precision under different recalls, where precision is how many of the retrieved items are accurate, and recall is how many of all accurate items are retrieved. In this work, to represent the accuracy of the proposed crop-row detection framework, we use the average of APs for Intersection over Union(IoU) from 0.5 to 0.95 with a step size of 0.05.
Discussion
The contribution of this paper is to design a workflow based on UAV and remote sensing technology to estimate the emergence rate of early maize crops, and to mark the sparse regions to provide references for farmers to take corresponding measures. This solution solves the small coverage and low efficiency problems of using ground vehicles and insufficient spatial resolution problem of using satellite data. Compared with other UAV-based solutions that focus on identifying individual crops, our
Conclusion
In this work, we designed and implemented a system that uses state-of-the-art deep neural networks and computer vision algorithms for early-stage maize stand count and locating regions with poor emergence in the fields by utilizing geometric descriptor information extracted from UAV-based remote sensing imagery. The system was tested in two different locations and seasons for robustness evaluation with 95.8% accuracy for site 1 (30 meters flight altitude) and 82.7% accuracy for site 2 (45
CRediT authorship contribution statement
Yan Pang: Writing - original draft, Software. Yeyin Shi: Supervision, Project administration, Writing - review & editing, Conceptualization. Shancheng Gao: Software. Feng Jiang: Formal analysis. Arun-Narenthiran Veeranampalayam-Sivakumar: Data curation. Laura Thompson: Data curation, Resources, Writing - review & editing. Joe Luck: Data curation, Resources, Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (32)
- et al.
Crop/weed discrimination in perspective agronomic images
Comput. Electron. Agric.
(2008) - et al.
Automatic segmentation of relevant textures in agricultural images
Comput. Electron. Agric.
(2011) - et al.
Research of uav flight planning parameters
Positioning
(2012) - et al.
Automatic inter-plant spacing sensing at early growth stages using a 3d vision sensor
Comput. Electron. Agric.
(2012) - et al.
A semi-supervised system for weed mapping in sunflower crops using unmanned aerial vehicles and a crop row detection method
Appl. Soft Comput.
(2015) - et al.
Wheat yellow rust monitoring by learning from multispectral uav aerial imagery
Comput. Electron. Agric.
(2018) - et al.
Evidence-based design heuristics for idea generation
Des. Stud.
(2016) - et al.
Image understanding research for automatic target recognition
IEEE Aerosp. Electron. Syst. Mag.
(1993) - et al.
Blob detection by relaxation
IEEE Trans. Pattern Anal. Mach. Intell.
(1981) - et al.
Machine learning techniques for the assessment of citrus plant health using uav-based digital images
Using remote sensing to assess stand loss and defoliation in maize
Photogram. Eng. Remote Sens.
Deep residual learning for image recognition
Mask r-cnn
Mask scoring r-cnn
Cited by (0)
- 1
Co-senior authors who made equal contributions to manuscript.