| Table of Contents | ||||
|---|---|---|---|---|
|
“The scarcity of public image datasets remains a key bottleneck in developing next-generation computer vision and intelligent systems for precision agriculture.” (Lu and Young 2020)
Why?
Developing image datasets is difficult, especially for agricultural applications. Collecting and annotating images pose major challenges.
Image Collection
Data hungry deep learning models require hundreds of thousands, sometimes millions, of images to train and become invariant to changes in location, lighting, backgrounds, or other factors. Devising high-throughput systems is necessary to capture hundreds of thousands of images of weeds at various growth stages while affected by field-like conditions. A network of agronomic, hardware, and software engineers is needed to automate large scale image collection.
Labeling is Time Consuming
Pixel-wise labels must be accurate. The process of labeling by hand can take minutes to hours even when using third-part annotation tools. Labels for semantic segmentation are the most time consuming. The need for high accuracy, complex leaf structures, and amorphous shapes make labeling plants one of the most difficult labeling tasks. Many have reported the high time requirement needed for labeling images of weeds.
...
Source
...
Annotation Technique
...
Scene type
...
Time
...
...
manual segmentation of cutouts
...
Simple
...
5-30 min / real image
...
Skovsen et al. (2019) (in conversation)
...
manual segmentation of cutouts
...
complex (field)
...
hour(s) / real images
...
...
manual segmentation of drone images of 3 classes
...
complex (field)
...
60 min / image
...
...
manual segmentation 3 classes
...
complex (field)
...
2-4 hrs / image
Labeling is Expensive
Texas A&M Case Study
Used a third-party labeling service to sematically label images of agricultural weeds and plants at early growth stages roughly 2-6 weeks after emergence. Details are as follows:
Company: Precise BPO Solution
1000 images
11726 segments
$0.125 per segment*
*While TAMU was given a discount of $0.095 per segment, $0.125, the non-discounted price, is used here. It is unlikely the company will provide the same discount for images over 1000.
Time to label all images - 2.5 weeks
Number of workers unknown
$1465.75 total cost
Other Labeling Services
Precise BPO Solutions were relatively inexpensive compared to other more known labeling services. However, their total time (2.5 weeks for 11,726 segments) is not scalable. Using more workers may decrease turn-around-time but brings increased costs.
Using third-party labeling services like Google AI Platform and Amazon SageMaker come with high costs when considering scale and time. Less time for labeling requires more workers which increases costs.
Google Cloud (AI Platform)
uses “unit” pricing. For example, 2 segments x 2 workers = 4 units.
Prices start at $870 for 1,000 units
Image segmentation is their most expensive labeling task
Amazon SageMaker (Mechanical Turk)
Pricing for number of reviewable objects, in our case segments, plus tasks
$0.08 per object review + $0.84 per semantic labeling task
reviewing 16272 per week * (730 hours in a month / 168 hours in a week) = 70705.71 per month
semantic segmentation is most expensive labeling task
Amazon SageMaker (vendor: Cogito)
highest rated labeling vendor in amazon marketplace
...
Company
...
Workers
...
Segments
...
Worker hours
...
Cost
...
...
1
...
1000
...
$870
...
Amazon SageMaker (Mechanical Turk)
...
1
...
1
...
$0.08 + $0.84
...
Amazon SageMaker (vendor: Cogito)
...
1
...
$5.04
...
...
3,600 hrs or 360,000 annotations
...
$5400 / year
Agricultural Scenes are Diverse
High individual diversity from differences in growth stages and effects of biotic and abiotic factors on morphology
Large intra-species variation and similarity with other species
Complex agricultural scenes as the result of effect from climate, soil, ground residue, and other weed populations.
Labeling Cost Projections for SemiField
Precise BPO Solutions
We can estimate the expected costs and time for SemiField data collection using the Texas A&M numbers.
here we use a smaller average of 8 segments per image (instead of 11.726 like TAMU)
$0.125 per segment
2.5 weeks (13 working days from 8am - 5pm) = ~104 worker hours
104 hours translates to 32 seconds per segment (104 / 11726)
...
Images
...
Segments
...
Worker Hours estimate
...
BPO Solutions
...
Google Cloud AI Platform**
...
Amazon Turk**
...
1,000*
...
11,726
...
6,249.958
...
$1,465.750
...
$20,403
...
$20,638
...
1,000
...
8,000
...
4,264
...
$1,000
...
$13,920
...
$14,080
...
10,000
...
80,000
...
42,640
...
$10,000
...
$139,200
...
$140,800
...
25,000
...
200,000
...
106,600
...
$25,000
...
$348,000
...
$352,000
...
50,000
...
400,000
...
213,200
...
$50,000
...
$696,000
...
$704,000
...
100,000
...
800,000
...
426,400
...
$100,000
...
$1,392,000
...
$1,408,000
...
250,000
...
2,000,000
...
1,066,000
...
$250,000
...
$3,480,000
...
$3,520,000
*TAMU example **Using 2 workers for quicker turn around timeWhile the potential of computer vision and artificial intelligence to revolutionize precision agriculture is immense, the lack of accessible and comprehensive image datasets is a significant roadblock. Why is this the case?
The Challenges of Image Collection and Annotation
Image Collection is Demanding:
Deep learning models thrive on massive amounts of data, often requiring hundreds of thousands or even millions of images to learn effectively. This demands high-throughput systems that can efficiently capture the diversity of agricultural scenes, from various growth stages to varying field conditions.
Creating such systems requires a multidisciplinary effort, involving agronomists, hardware engineers, and software developers.
Agricultural Scenes are Inherently Diverse:
Plants undergo significant changes in appearance throughout their growth cycle.
Environmental factors like pests, diseases, and weather can drastically alter plant morphology.
Weeds present a particular challenge due to both high intra-species (within a species) and inter-species (between species) variations.
Agricultural fields are complex environments, influenced by climate, soil, ground residue, and the presence of other plant species.
Labeling is a Time-Consuming and Expensive Bottleneck:
Pixel-wise annotations, essential for training segmentation models, are especially labor-intensive. Even with specialized tools, manually labeling a single image can take minutes to hours.
The need for high accuracy, coupled with the complexity of plant structures, makes labeling agricultural images a particularly demanding task.
Numerous studies have documented the significant time investment required for labeling weed images.
High Costs: Third-party labeling services, while offering a solution, can be prohibitively expensive, especially for large datasets. The Texas A&M case study illustrates this, with costs exceeding $1400 for labeling just 1000 images.
AgIR: A Solution to the Data Challenge
The Agricultural Image Repository (AgIR) addresses these challenges by providing a comprehensive, well-annotated dataset that combines:
High-throughput Semi-Field Data: Collected under semi-controlled conditions for scalable and automated annotation.
Real-World Field Data: Capturing the complexity and diversity of agricultural environments.
By combining these two data sources and utilizing innovative annotation strategies, AgIR aims to accelerate research and development in precision agriculture, ultimately leading to more efficient, sustainable, and productive farming practices.