KaraAgroAI commited on
Commit
74df710
1 Parent(s): 469f0b4

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +82 -0
README.md CHANGED
@@ -1,3 +1,85 @@
1
  ---
2
  license: cc-by-4.0
 
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: cc-by-4.0
3
+ datasets:
4
+ - KaraAgroAI/KaraAgroAI/Drone-based-Agricultural-Dataset-for-Crop-Yield-Estimation
5
+ language:
6
+ - en
7
+ library_name: yolo
8
+ tags:
9
+ - object detection
10
+ - vision
11
+ - yolo
12
+ pipeline_tag: object-detection
13
+ metrics:
14
+ - mape
15
  ---
16
+
17
+ ## Drone-based Agricultural Dataset for Crop Yield Estimation
18
+
19
+
20
+ # Table of Contents
21
+
22
+ 1. [Accomplishments and lessons learned](#accomplishments-and-lessons-learned)
23
+ 1.1 [Summary of activities and accomplishment](#summary-of-activities-and-accomplishment)
24
+ 1.2 [Key insights gained](#key-insights-gained)
25
+ 2. [The final dataset and other deliverables](#the-final-dataset-and-other-deliverables)
26
+ 3. [Sustainability and use of the dataset](#sustainability-and-use-of-the-dataset)
27
+ 4. [Impact, storytelling and publications](#impact-storytelling-and-publications)
28
+
29
+ ## Accomplishments and lessons learned
30
+
31
+ ### Summary of activities and accomplishment
32
+
33
+ KaraAgro AI Foundation
34
+ During the reporting period, the following activities were conducted:
35
+ Meeting with partners and stakeholders to organize data collection exercises.
36
+ Travelled to cashew and cocoa farms in the Bono and Eastern region respectively
37
+ Multiple trips to the Bono Region and Eastern Region in Ghana were required to reach the target number of images for cashew and cocoa crops. The collection was done on multiple farms.
38
+ Collected image data of cashew and cocoa crops using a DJI P4 Multispectral Drone
39
+ Collected 4,715 instances of cashew images and 4,069 instances of cocoa images
40
+ Annotation of cashew trees, flowers, immature, mature, ripped and spoilt cashew and cocoa fruits was done over a period of 2 months. This involved three annotators with supervision from an agricultural scientist. Before annotation, an annotation protocol was designed by the agricultural scientist for annotators to follow. The tool used for the annotation was Makesense.ai
41
+ Recorded quantitative measures of progress, including the number of observations and recordings collected. Every important detail relating to the data collection has been recorded and made available.
42
+
43
+ Makerere AI Lab
44
+ During the project period between July 2022 to July 2023, the Uganda team carried out several activities to fulfil project objectives. The activities that were undertaken included: collaboration meetings with partners and stakeholders, pilot field work, data collection, training workshops, and data annotation and validation exercises. The research team held a number of planning and project work meetings with our partner organizations, namely the National Coffee Research Institute (NaCORI), and with the leadership of the Uganda Cashew Farmers Network. NaCORI is an agricultural research institute under the National Agricultural Research Organisation (NARO) that is mandated to generate, develop and promote technologies, methods and knowledge to improve the production and productivity of coffee and cocoa in Uganda. These meetings resulted in signing a Memorandum of Understanding with NaCORI for research collaboration and setting up joint fieldwork teams. The pilot field works were carried out to achieve the following objectives: 1) pre-test guidelines for drone image data collection, 2) pre-test guidelines for ground truth coffee yield data acquisition, and 3) test drone equipment and evaluate the quality of images captured. These activities were carried out on coffee farms owned by NaCORI, one of our partners in this project. These pilot activities were beneficial in planning field data collection activities and refining our data collection guidelines and protocols.
45
+ We carried out field data collection activities in Uganda's central, southern, eastern, and northern regions to ensure that our datasets were geographically representative regarding coffee and cashew growing areas. The main coffee varieties in Uganda, namely Arabica and Robusta, were represented in our dataset. A total of 6,086 drone images, comprising 3,000 for coffee and 3,086 for cashew, were collected due to these field activities. A total of 3000 coffee yield data points were collected. Additionally, we also carried out field data collection activities in March 2023 from various prominent cashew nut growing districts in the Eastern region, North Eastern region and Central region Uganda. In this data collection, we collected 3086 images of cashew fruits (apples) and nuts at 4 stages of maturity (premature, unripe, ripe, spoilt) and flowers.
46
+ The images were captured for coffee and cashew using an unmanned aerial vehicle, commonly known as a drone, which is a small aircraft without any human pilot or crew. The drone we used in particular had an inbuilt high-quality camera and was controlled by an experienced drone pilot to take high quality images of the coffee and cashew tree foliage. These images were captured according to the imaging guidelines created for the data collection activities.
47
+ We created annotation protocols for the coffee and cashew data annotation and assembled teams that undertook the task of annotation and validation of the datasets. For the coffee dataset, we had a team of five annotators that annotated the 3000 coffee images using the VIA annotation tool. The team was guided by the coffee annotation protocol which highlights how to do the annotation and the labels for the coffee cherries. The annotation of the cashew dataset involved a team of two annotators and one validator. The team annotated and validated 3086 cashew images using the makesense ai tool. The dataset will be made publicly available through Hugging Face. Hugging Face is a platform that allows users to share machine learning models and datasets. The dataset will be published in August 2023 and can be accessed through a DOI: doi:10.57967/hf/0941
48
+
49
+ How did your project fill a data gap?
50
+ The gap that our project addressed is the lack of machine learning datasets for crop yield estimation, especially for coffee and cashew, which are important sources of livelihood for millions of households in Sub-Saharan Africa. Yield estimation provides farmers with an opportunity to make good business decisions and appropriately plan towards their equipment, fuel and labor needs, ensure they have enough storage available, cash-flow budgeting, and make early marketing decisions. Timely harvests also provide farmers the opportunity to ensure healthy and fresh produce, and therefore better sales and income. Conventional methods of yield estimation are expensive, require a lot of labor, take time and are prone to large errors due to incomplete ground observations, which leads to poor crop yield estimations, affecting the ability of farmers to appropriately plan and manage their fields and production pipelines. However, development of accurate machine learning methods for crop yield estimation is dependent upon availability of suitable datasets. Our project therefore, filled this data gap for coffee and cashew by creating high quality image datasets for these two crops. The datasets we created have many potential use cases including fruit detection, counting, and for classifying the different stages of fruit development. In addition, the foliage images may be used for crop variety identification. Coffee yield estimation may be based on the cherry as well as foliage images using deep learning techniques.
51
+
52
+
53
+ ### Key insights gained
54
+
55
+ KaraAgro AI Foundation
56
+ Standardized data collection protocols are crucial for ensuring reproducibility and comparability of results across different experimental settings
57
+ Class imbalances in the datasets are a representation of real life and it is difficult to obtain balanced datasets
58
+ Collecting agricultural data is made complicated by the effect the environment/season has on the maturity stage of crops
59
+ Makerere AI Lab
60
+ The Uganda team learned some useful lessons while undertaking this project, which include; Collaboration with partners and engaging with stakeholders is key in achieving project objectives. In our experience, it was important that we worked closely with our partner organization the NaCORI right from the beginning by onboarding their team, through to setting up fieldwork teams and planning field activities, to training workshops and actual data collection, annotation, and validation. We also learned that it is important to carefully plan and synchronize the timing and scheduling of fieldwork activities for data collection with the desired development stage of the target crop. For example, suppose the activity is to do with crop yield data collection. In that case, it is important to plan and schedule fieldwork activities several weeks ahead of the beginning of the harvest season. Fieldwork plans, data collection guidelines, protocols, and imaging equipment (e.g., UAV/drone equipment) need to be pre-tested to ensure that they work as planned and give the desired results in terms of image quality prior to carrying out actual data collection fieldwork.
61
+
62
+
63
+ ## The final dataset and other deliverables
64
+
65
+ KaraAgro AI Foundation
66
+ Yes, we have completed the collection and annotation of the dataset. For cashew, we collected 4715 images and annotated 6 classes, namely: flower, immature, mature, ripe, spoilt, and cashew_tree. For cocoa, we collected a total of 4,069 images and annotated the same 6 classes annotated for cashew. Each drone image has a dimension of 1,600 by 1,300 pixels and contains metadata including timestamp and geographic location of an image. Images were captured from multiple angles including aerial shots and side shots.
67
+ Data for cashew and The dataset will be maintained by the karaAgro AI team. Further discussion will be had concerning the expansion of the dataset.
68
+ Makerere AI Lab
69
+ We completed the collection and annotation of coffee and cashew image datasets as specified in our original project proposal. For coffee, we have collected and annotated a total of 3,000 images as planned for in the original project proposal. Annotated objects of interest include coffee cherries (fruit) at three stages of development namely ripe, unripe, and spoilt. A total of 3,086 cashew images were collected, which exceeded the planned target of 3,000 images. Annotated objects of interest include cashew apples (fruit) and flowers. The fruits belong to four classes namely premature, unripe, ripe, and spoiled. All images are in both jpeg and geotiff file formats. The annotation file format is Pascal VOC XML.
70
+ Each drone image is approx. 5 megabytes in size 4,000 by 3,000 pixels in dimension and 72 pixels/inch in Dots per inch (DPI). Each image contains multiple bounding boxes describing the regions of interest as stated in the previous paragraph. Each image also contains metadata including information on the date the image was captured (timestamp) and geographic location (latitude and longitude) where it was captured.
71
+ For coffee, each image is a representation of a coffee plant from either the top or from opposite lateral sides. For cashew, images were captured from different lateral sides.
72
+ The imaging equipment consisted of a DJI Mini 3 pro drone equipped with a high quality camera that had a 48 MP 1/1.3 inch CMOS sensor, lens with aperture of f/1.7 and focus range of 1m to ∞,shutter speed of 2-1/8000s and an ISO range of 100-6400 (Auto and Manual)
73
+ During the annotation process, we used two annotation tools; VIA and Makesense AI. We used VIA for annotation of the coffee dataset. VGG Image Annotator (VIA) is a simple and standalone manual annotation software for image, audio and video.VIA is an image annotation tool that can be used to define regions in an image and create textual descriptions of those regions. VIA runs in a web browser and does not require any installation or setup. For annotation of the cashew dataset, we used the Makesense AI tool which is a free, open source tool for labeling images. Makesense Ai is an AI-powered annotation tool that allows developers, data scientists, and researchers to quickly and accurately label and annotate datasets for machine learning and computer vision projects. The tool provides an intuitive user interface for labeling, configuring, and managing image processing tasks. It provides powerful tools for annotation and image manipulation, as well as a variety of features to help streamline the annotation process
74
+
75
+ Our datasets comply with the Findable, Accessible, InterOperable, and Reusable (FAIR) data principles. The datasets along with associated metadata (datasheets, annotation protocols, etc) have been uploaded to DataVerse, an open-source data repository. The datasets have been assigned a Digital Object Identifier (DOI), a permanent unique identifier to facilitate findability and accessibility. In addition, the dataset have been mirrored to local servers as a form of backup. The metadata is citable and includes domain-specific and file-level data that map to metadata standards within machine learning, computer vision, data analysis - geospatial and time series analysis to make it Interoperable. The metadata has been published and made available to provide a description of the datasets, data acquisition, preprocessing, and annotation procedures, envisaged use cases for our dataset, and any other information that supports understanding the context and composition of the data and ensure that they are reusable. Our datasets along with their associated metadata may be accessed and downloaded via this link : doi:10.57967/hf/0941
76
+
77
+ Our dataset has been published under the Creative Commons Attribution 4.0 International Public License (CC-BY 4.0). This licence gives anyone permission to use, copy, edit, transform and redistribute the dataset as they wish for any purpose, including use for commercial purposes. However, the user of this dataset is required to give appropriate credit by citing us as the source of the original dataset. An appropriate method for citation of this dataset has also been provided.
78
+
79
+
80
+ ## Sustainability and use of the dataset
81
+
82
+ The dataset could be used for further research including crop abnormality detection. The machine learning data community is a potential user of the dataset.
83
+
84
+
85
+