An instance segmentation dataset of cabbages over the whole growing season for UAV imagery

Crop growth monitoring is essential for both crop and supply chain management. Conventional manual sampling is not feasible for assessing the spatial variability of crop growth within an entire field or across all fields. Meanwhile, UAV-based remote sensing enables the efficient and nondestructive investigation of crop growth. A variety of crop-specific training image datasets are needed to detect crops from UAV imagery using a deep learning model. Specifically, the training dataset of cabbage is limited. This data article includes annotated cabbage images in the fields to recognize cabbages using machine learning models. This dataset contains 458 images with 17,621 annotated cabbages. Image sizes are approximately 500 to 1000 pixel squares. Since these cabbage images were collected from different cultivars during the whole growing season over the years, deep learning models trained with this dataset will be able to recognize a wide variety of cabbage shapes. In the future, this dataset can be used not only in UAVs but also in land-based robot applications for crop sensing or associated plant-specific management.

a b s t r a c t Crop growth monitoring is essential for both crop and supply chain management.Conventional manual sampling is not feasible for assessing the spatial variability of crop growth within an entire field or across all fields.Meanwhile, UAVbased remote sensing enables the efficient and nondestructive investigation of crop growth.A variety of crop-specific training image datasets are needed to detect crops from UAV imagery using a deep learning model.Specifically, the training dataset of cabbage is limited.This data article includes annotated cabbage images in the fields to recognize cabbages using machine learning models.This dataset contains 458 images with 17,621 annotated cabbages.Image sizes are approximately 500 to 10 0 0 pixel squares.Since these cabbage images were collected from different cultivars during the whole growing season over the years, deep learning models trained with this dataset will be able to recognize a wide variety of cabbage shapes.In the future, this dataset can be used not only in UAVs but also in land-based robot applications for crop sensing or associated plant-specific management.
© 2024 The Author(s

Value of the Data
• This dataset was created by object-based annotation of individual cabbage with laborious manual effort s.The use of distinct dat aset and inst ance segment ation models such as Mask R-CNN [ 1 ] and YOLACT [ 2 ] enables on-farm assessment of individual cabbage growth (e.g., leaf area index and biomass) by quantifying individual cabbage contours.• This dataset can also be used for object detection such as YOLO [ 3 ] because the JSON file includes the bounding box data if practitioners only need to count the number of cabbages.• The dataset is highly complementary to the existing dataset [ 4 , 5 ] because it consists of multiple cultivars with different morphological traits, and the images were taken during the whole growing season.Since the image data is classified by cultivar, location and image acquisition time, it is easy to select a specific cultivar or growing season depending on the purpose of model optimisation and validation.• The dataset can be used for actual crop monitoring.In the future, it is expected to enable automatic plant-level management coupled with land-based agricultural robots.

Data Description
• (a) Image number • (b) The day the image was taken • (c) Location All images have annotated information in a JSON file "annotated.json".Fig. 1 shows the segmentation masks of the dataset.Each cabbage mask was enclosed along the outline and saved as the label name "cabbage".

Field data collection
The RGB images of cabbages were collected in 2020, 2021, and 2022 in farmers' fields in three cities, including Yoro, Sunomata and Kaizu, Gifu Prefecture, Japan.Table 1 shows the cul- tivars grown in each city and transplanting day.In Kaizu on 2020 and 2021, cabbages were cultivated by transplanting one row in a ridge.Others were cultivated by transplanting two rows in a ridge.This also makes images look different due to the different plant density and transplant spacing.Fertilisation and pest/disease management was performed according to the recommendation suggested by the local crop advisory service.All the cabbages were grown under rainfed conditions.During the cabbage growing season from September to December, RGB images were taken using a UAV.All images were captured between 9:0 0 and 16:0 0 under various cloud conditions from clear sky to thick clouds.Given the temporally varying light intensity even within a day and varying sun angles over the growing seasons from summer to winter, entire dataset covers the effect of a wide range of varying illumination conditions on image quality.In 2020, Phantom4 (DJI, Shenzhen, China) was used for image acquisition at a 20 m altitude with 75% front and side overlap.In 2021 and 2022, the digital camera α6600 (Sony, Tokyo, Japan) mounted on MATRICE300 (DJI, Shenzhen, China) was used for image acquisition at a 30 m altitude with 70-80% front and side overlap.The coordinates of RGB images were measured using KlauPPK (Klau Geomatics, New South Wales, Australia) with a 0.03-m accuracy.

Data preprocessing
The orthomosaic process was performed using Pix4D mapper version 4.6.4(Pix4D, Prilly, Switzerland) based on the processing template 3D Maps.The orthomosaic images were split into 515-10 0 0 pixel squares using GDAL as a Python module.The spatial resolution of the resultant split images ranged from 3.3 to 6.3 mm.A total of 458 split images were randomly selected from the split images.The cabbage masks were manually annotated using the COCO annotator [ 6 ].The drawing tablet Cintiq 16 (Wacom Co., Ltd, Saitama, Japan) was used for the accurate and efficient annotations.To accurately draw the semantic annotations for the individual cabbage, the application of expert judgement derived from experiences in field survey becomes indispensable particularly in situations involving the overlapping of multiple cabbages.Therefore, highly trained technician was involved in the annotation process, and every annotation were carefully checked by the authors.

Limitations
None.

Fig. 1 .
Fig. 1.Raw and annotated images of the dataset.

Table 1
Transplant day of each day and cultivar.