Size Estimation of Oyster Mushroom Clusters for Computer Vision-Enabled Mushroom Growth Monitoring

Automated growth tracking of non-symmetrical agricultural products in real-time has been a promising sector of smart farming. Pleurotus ostreatus, poses an exceptional challenge in the field of fruit detection and growth monitoring, due to its intricate morphology and the spatial variability of the clusters within the crop. The main challenges in the field comprise of the lack of an extended annotated image dataset that can serve as a reference method and transfer learning for specialized computer vision models, as well as lack of monitoring during different stages of maturity. In this work, an experimental method for size estimation of oyster mushrooms was developed, tested and validated, with main objective to assist computer vision models through providing robust empirical relationships between real dimensions and pixel-generated measurements. Based on the development of an experimental methodology including accurate manual measurements, the study was able to provide comprehensive raw and annotated datasets of static and dynamic images, as well as establishing empirical relationships between real-size and pixels, using Python code. The experimental process entailed the acquisition of size and data for 56 oyster mushroom clusters in a farm in Hungary, cross-validated with 24 RGB images captured with a RealSense camera and a common DSLR camera. Additional experiments with the same equipment involved size and measurements on grey oyster mushroom substrate blocks in the lab during all phases of mushroom growth, under controlled illumination, temperature and humidity. The image dataset consisting of over thousand measurements and their corresponding computer-vision measurements in pixels, was annotated with Computer Vision Annotation Tool (CVAT). The ground truth estimations extracted from the field measurements were adequately estimated the pixel estimations provided by the computer-vision masks. Finally, a pixel-wise estimation of the area of the cluster was drawn and expressed through growth curves during the mushroom growth cycle. Although the model underperformed in the complex environment of the farm, in the laboratory conditions it presented R2 of 0.80 (SEE =1.97 cm), 0.85 (SEE=1.53 cm) and 0.96 (SEE=1.59 cm), for front, top and side camera view, respectively. The aforementioned dataset will serve as a valuable training material and a robust reference basis for the development of more efficient detection and yield forecasting not only in the mushroom field, but also for more irregularly shaped agricultural products, like the oyster mushrooms. In the future, the camera specifications as well as the establishment of a reference plane and the use of photogrammetry would further expand the findings of this study.