A soybean seedlings dataset for soil condition and genotype classification

We have witnessed the exponential growth of the number of Deep Learning techniques as well as their technical and theoretical advancements. Nowadays, the researchers’ effort is also focusing on the applicability of those well-founded concepts. One target is precision agriculture, the area of farming management guided by observation, measurement, and data analysis. Generalization is a real challenge regarding the farm environment. The visual information used during the model’s test phase may be exposed to a significant variation in light and weather conditions and visual differences due to plant growth stages. Another challenge is the cost of collecting and labeling data. Convolutional Neural Networks (CNNs) have been used in many applications, such as detecting diseases, distinguishing animal species and plant genotypes, and identifying the plant’s environmental characteristics. A large volume of data is required to enable the model to generalize well in these real farm conditions. One open problem is the study of which soybean genotypes best adapt to the soil compaction condition, aiming to increase crop productivity. This problem is a candidate to be solved using CNNs; however, an image dataset is necessary to train the model. In this paper, we proposed a dataset with approximately 1,000 soybean seedlings images from 30 different genotypes grown in compacted and non-compacted soil conditions. To demonstrate the dataset representativeness, we trained classic CNN models to classify soil conditions based only on the visual information and then interpreted the model learning using visual explanation methods. We also raised possible future work for using the dataset in the precision agriculture field.