The MiPlo Datasets

| Paper | Code |

The Mini Plot (MiPlo) datasets are large-scale RGB image datasets consisting of high-resolution images with multiple types of nutrient treatments annotated by agronomy experts. It consists of the Mini Plot Barley (MiPlo-B) dataset and the Mini Plot Winter Wheat (MiPlo-WW) dataset.

Statistics

Dataset	#Images	#Class	Dates	#Views	Year
MiPlo-B (Barley)	18559	6	16	20	2022
MiPlo-WW (Winter Wheat)	12466	6	13	20	2023

The Mini Plot Barley (MiPlo-B) dataset consists of 18559 images with 6 nutrient treatments (-N, -P, -K, -B, -S, ctrl) annotated, ranging from 21.06.2022 - 20.07.2022 (16 dates). It contains two genotypes: Barke (9305 images) and Hanna (9254 images). For each genotype, each treatment was repeated 4 times, resulting in 24 containers, each of which has a unique ID. Six unique containers with six different nutrient treatments were selected as the test set while the other containers as the training set (#train:#test$\approx$75%:25%).
The Mini Plot Winter Wheat (MiPlo-WW) dataset has 12466 images with 6 treatments (-N, -P, -K, -B, -S, ctrl) annotated, ranging from 12.05.2023 - 24.05.2023 (13 dates). It contains two genotypes: Julius (6253 images) and Meister (6213 images). The ID settings are the same as above.

Experimental Setup

The MiPlot experiments enable controlled trials of nutrient deficiencies under simulated field conditions. The crops (two genotypes) were grown in mineral soil in containers (“Big Box”, L x B x H: 111 x 71 x 61 cm, vol.: 535L) and sown in rows in a density according to agricultural practice. The soil was a mixture of nutrient-poor loess from a 5 meter depth of an opencast pit mine and quartz sand 0.5 – 1 mm. To expose the plants to environmental factors, e.g., wind, radiation, and precipitation, the containers were positioned in an outdoor area and equipped with a fertigation system of porous drip pipes to allow additional water supply and individual fertilization with nutrient solutions per container.
To transfer soil microorganisms to the experiments, the containers were inoculated with soil slurry from the non-fertilized plot of a long-term fertilizer-field experiment. For each genotype, the containers were placed in three rows of ten containers each on a leveled concrete platform. The 30 containers were divided into seven treatments (ctrl, -N, -P, -K, -B, -S, unfertilized) with four replicates each, as well as two additional containers for invasive investigations, in a randomized block design. In this repository, 24 containers with six nutrient treatments (ctrl, -N, -P, -K, -B, -S) were selected for evaluation, because the containers with unfertilized treatment showed distinct patterns, i.e., only pipes.

Image Acquisition Protocol

The RGB images in the MiPlo datasets were taken from 24 containers, each of which was subjected to a type of nutrient treatment. All of the images with the size of 7296 x 5472 were captured by a Huawei P20 Pro smartphone with a triple camera from Leica under natural lighting conditions. Specifically, the images were taken under different conditions in terms of height, viewpoint, light, and weather to reflect realistic conditions. As a result, crops within each container have been captured over the growing season multiple times and each time from multiple views (20 views on average). The images were annotated by the date, genotype, and six nutrient treatments (ctrl, -N, -P, -K, -S, -B), where ``-'' stands for the omission of the corresponding nutrient (N: nitrogen, P: phosphorous, K: potassium, B: Boron, S:Sulfur). Plants in the control group `ctrl' do not suffer from nutrient deficiencies.

✏️ Citation

If the MiPlo datasets are helpful for your research, please consider giving a star ⭐ and citation 📝.

@inproceedings{yi2024mvmatch,
    title={MV-Match: Multi-View Matching for Domain-Adaptive Identification of Plant Nutrient Deficiencies},
    author={Yi, Jinhui and Luo, Yanan and Deichmann, Marion and Schaaf, Gabriel and Gall, Juergen},
    booktitle={BMVC},
    year={2024}
}