Dataset of feed bunk score images of cattle feedlot

Bunk management is an important technique to minimize the variations in consumption in feedlot cattle and can be performed according to the South Dakota State University classification system. The use of information and communication technology (ICT) can help, in an objective way, in the interpretation of these measurements. We created a dataset with the objective to develop an automatic classification method of feed bunk score. In May, September and October on the 2021 and September on the 2022 we captured 1511 images in the morning on the farms, in natural lighting conditions with different angles and backgrounds and at a height of about 1.5 m from the bunk. After acquisition data, each image was classified according to its score classification. Additionally, we resized the images to 500 × 500 pixels, generated annotations files, and organized the dataset in folders. The images in this dataset can be used to train and validate a machine learning model to classify feed bunk images. This model can be used to develop an application to support bunk management.


a b s t r a c t
Bunk management is an important technique to minimize the variations in consumption in feedlot cattle and can be performed according to the South Dakota State University classification system. The use of information and communication technology (ICT) can help, in an objective way, in the interpretation of these measurements. We created a dataset with the objective to develop an automatic classification method of feed bunk score. In May, September and October on the 2021 and September on the 2022 we captured 1511 images in the morning on the farms, in natural lighting conditions with different angles and backgrounds and at a height of about 1.5 m from the bunk. After acquisition data, each image was classified according to its score classification. Additionally, we resized the images to 500 × 500 pixels, generated annotations files, and organized the dataset in folders . The images in this dataset can be used to train and validate a machine learning model to classify feed bunk images. This model can be used to develop an application to support bunk management.
© 2023 The Author(s

Value of the Data
• This dataset can be useful to develop deep learning algorithms to classify cattle feed images in relation to the amount of leftovers, composition and feed bunk attributes. • The set of images of leftover feed in bunk can contribute to the development of research on new methods of managing feedlots for confined cattle. • The methods developed and technological tools obtained through this data can enable farmers and bunk readers to support the management of the animal nutrition and thus optimize the farms' profits. • This dataset can be used to explore the potentialities of the applications of computer vision in livestock farming operations. • This dataset has the potential to develop new tools capable of assisting in the balancing of rations, in the recognition by images of the ingredients used, in the observations of the animals' preference for some food, and reduce costs with cattle feeding.

Objective
This dataset was created to develop an automatic classification method of feed bunk score. The images could be used to train and validate a machine learning model. This model could be embedded in an application, for mobile devices. In this application the feeder would capture the images using the camera and the application would suggest the adjustments.

Data Description
The Feed Bunk Score Images -FBSI dataset is a set of images of leftover rations on feed bunks in feedlot cattle. These images are used for reading and classification of the bunk score. This technique is a visual assessment and consists of analyzing the amount of feed left in the bunk over the last 24 h. For each volume of leftovers, specific scores are assigned suggesting adjustments in the amount of feed to be provided during the day, according to the needs of the animals in each herd [1][2] . In this system, the animals' residual feed is classified into six bunk scores according to the South Dakota State University (SDSU) score system. For scores 0 and 1/2 it is recommended an increment adjustment, for score 1 a maintenance adjustment, and for scores 2, 3 and 4 a decrement adjustment, according to Table 1 [3] . This visual assessment avoids wasting feed. Since among the operational expenses of the feedlot, feed costs can reach 60% to 80% of the total, avoiding wasting feed is essential for the producers [4] .
The dataset is composed of 1511 images of 6 different scores, distributed according to Table 2 . These images represent different diet compositions adopted on the farms. Furthermore, in this dataset, the feed bunks differ in terms of their material and model. Most feed bunks are made of cement, but some are made of wood with the background covered in rubber or plastic material. Fig. 1 shows a sample of each score image. Because of the high similarity between classes and high intra-class variability, distinguishing some classes can be a complex task. For these reasons, the images labeling in our dataset was performed by specialists in the animal nutrition field. where: • Score: represents the score according to Feed Bunk Scoring SDSU system (South Dakota State University). We also provide an XML file for each image containing some annotations. Table 3 presents a description and samples of the annotated attributes. For some farms, we could not collect the diet composition data. We indicated this field as "MD" (Missing Data) for these images. Since score 0 represents empty feed bunk, we defined the composition diet field in images of score 0 as "NApp" (Not Applicable). Additionally, we provide a CSV file named "annotations.csv" containing the annotation of all images. Table 1 Feed bunk scoring SDSU (South Dakota State University).

Score
Features Adjustment 0 There is no feed in the feed bunk Increase 1 kg of DM in the diet or 5% ½ Spread food -5% leftover Increase 1 kg of DM in the diet or 5% 1 Thin layer ( < 5 cm) -5 to 10% leftover Keep quantity provided 2 Medium layer (5 to 8 cm) -25% leftover Decrease 1 kg of DM in the diet or 5% 3 Thick layer ( > 8% cm) -50% leftover Decrease 1 kg of DM in the diet or 5% -note other problems 4 Intact food Decrease 1 kg of DM in the diet or 5% -note other problems   To standardize the comparison between classification methods and mitigate possible biases we suggest a 5-5-fold stratified cross-validation distribution for our dataset. For each distribution and each fold we provide CSV files containing the list of train, test and validation images.
For the purpose of the dataset, we do not provide a bounding box for each image. This is a pre-processing step, which can be performed or not, according to the pipeline designed by everyone who is going to explore the dataset. After acquisition data each image was classified according to its score classification. 3 Images processing and dataset organization October and November 2022 Images were resized to 50 0 × 50 0 pixels, XML and CSV files were generated and the dataset organized in folders.

Experimental Design, Materials and Methods
The feed bunk images were acquired using a high-resolution rear camera of iPhone 11 (Apple) and a digital camera Cyber-shot DSC -H300 (Sonny). We also recorded some videos and extracted the frames through Python scripts. To ensure variability, we sampled the videos every 30 frames. The images were collected on different days, between 6 am and 10 am, from eight farms in the state of Minas Gerais, Brazil. We performed the collections in May, September, and October 2021, and September 2022.
The original images had six different spatial resolutions (5152 × 3864, 4608 × 3456, 4032 × 3024, 1280 × 958, 1920 × 1080, and 960 × 1280 pixels) and sRGB color space. We reduced the images to 500 × 500 pixels, which is a suitable resolution for extraction of predictive visual features. Considering the state-of-the-art, several deep convolutional neural networks have input size of 224 × 224 pixels [5][6][7] . Another aspect to downscaling image resolutions is the dataset storage scalability, which can become a limitation if dataset updates with addition of new images occur.
To account for variations in environmental conditions, we manually captured the images in natural lighting conditions with different angles and backgrounds and at a height of about 1.5 m from the bunk. In each bunk visited the photos were captured across the entire length. We did not use a static camera that always points to the bunker, but the capture device stood in the hand of the person responsible for the capture, which pointed to the bunker. We collected the images during the daily routine of each farm. Therefore, they are representations of different bunk management and diet composition, and animals of different breeds whose feed consumption is not the same. Thus, this dataset is unbalanced since some scores are more commonly encountered than others. After acquisition data, we classified the images according to feed bunk scoring SDSU [4] and organized them into their respective folders. Table 4 summarizes the steps of data acquisition.

Ethics Statements
There is no conflict of interest. The data is available in the public domain.

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.