Avo-AirDB: An avocado UAV Database for agricultural image segmentation and classification

Unmanned aerial vehicles (UAVs) with on-board cameras have the advantage of providing Bird-view images (Aerial images). This type of image is considered as a rich source of information especially for intelligent agriculture. A dataset of 984 aerial images of avocado threes is made publicly available with a ground resolution of 2.7 cm per pixel. It has been collected from over 113 Hectares of an avocado farm in ALLAL TAZI region of Morocco using a DJI Phantom 4 Pro UAV. It comprises original bird view and annotated images. The dataset is available at https://data.mendeley.com/datasets/tvhh83r3hj/2


Specifications
Raw and analyzed Description of data collection: The dataset images were collected using the described UAV over a 113 Ha farm of Avocado trees. The camera angle was adjusted to 90 °vertically with the field. The speed and flight altitude were 9 m/s and 90 meters respectively. A 75% Longitudinal overlap was applied. Data source location: • City/Town/Region: Kenitra/Allal Tazi Region

Value of the Data
• The dataset represents a significant contribution to different applications in the digital agriculture field, such as: trees segmentation, trees counting and classification based on tree's crown surface, disease detection, etc. • The data can be used by artificial intelligent (AI) researchers in addition to agricultural researches and professionals. • The dataset is suitable for developing digital and precision agricultural systems.
• Collected data can be employed to train Artificial intelligence methods for image classification. • The presented data is the only public dataset of Avocado aerial high-resolution images in the African continent.

Objective
The progress of agricultural visual pattern recognition (especially for avocado) one of the fundamental aspects of human beings, has been relatively slow [1] . Due to the low number of countries with an avocado production of more than 50K tonnes per year (only 22 countries), what is causing a lack of relevant datasets to encourage the study of agricultural images of avocado and visual patterns with many unique characteristics. The objective of this Dataset is to encourage research on this challenging task.

Data Description
This paper describes a dataset of images collected by an unmanned aerial vehicle (UAV) from an avocado farm of more than 113 Ha leading to a set of 984 RGB images of 4864 × 3648 pixels. 93 images were annotated using the Make-sens.ai [2] and Apeer.com [3] platforms, forming Four classes, namely: Small, Medium, Large and background. In Fig. 1 samples from the developed dataset are provided while the Table 1 presents the folders and files organi-   zation of the dataset, and Table 2 and Table 3 represents the specifications of the used UAV and camera respectively. Fig. 2 shows the visible orthoimage using the Agisoft Metashape Software [4]

UAV flight mission and data acquisition
Flight planning ultimately helps achieve mission objectives, keep flight altitude restrictions in mind, avoid restricted airspace, and improve battery life performance. For cartography, it can be very helpful to plan the number of flight paths or waypoints, the time required to complete the flight mission, the area to be flown, the number of images taken in a given area and the overlap between the pictures. In addition, different parameters should be taken into account: Atmospheric conditions : Aerial survey missions are considered successful only if you obtain image quality that serves the scope of the planned mission. Weather conditions such as severe weather, crosswinds, or the wrong season can greatly affect the results. The position of the sun is a major factor in creating shadows. Generally, the best time to take aerial photos is between 10:00 am and 2:00 pm. Depending on the latitude of our scan area, around 12 o'clock is the best time due to the smallest shadows.
Weather : The weather is an essential parameter during the execution of the mission. For that, a forecast condition check in real-time is important. Table 4 gives the weather conditions during the mission time.
Planification : The mission planification necessity the setting up of different parameters for an optimal mission control. For our mission, we used DroneDeploy which is one of the most used and sophisticated Drone mapping software [5] . Table 5 presents the parameters of the acquisition mission.
The data collection was carried out using the described UAV ( Fig. 4. b) and based on the navigation scheme presented in Fig. 4. a, provided by Drone-deploy system in order to cover the entire avocado farm ( Fig. 3 ) and guaranteeing a maximum mission efficiency.

Data annotation
The collected data were manually annotated by our agricultural experts in the avocado field using two online platforms Apeer and makeSense. Images were annotated using a polygon outline, leading to Four classes in the first annotation (by Apeer platform): Small, Medium, Large  regions and the Background, and to three classes in the second annotation (by Make-sense platform): Small, Medium and Large trees. The aims are to propose annotated images that can be used in different machine learning applications. Table 6 summarized the statistics of the most related datasets of agriculture images. Seg refers to segmentation and Cls is refers to classification.

Ethics Statements
The dataset does not include animal in experiments, human subjects or data collected from social media.

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.

Data Availability
Avo-AirDB: An avocado UAV Database for agricultural image segmentation and classification (Original data) (Mendeley Data)