Machine vision based plant height estimation for protected crop facilities

The increasing demand for quality, year-round food production in limited space has led to the widespread adoption of protected cropping. Effectively monitoring and maintaining crops within these facilities requires substantial labour and expertise. Traditional manual monitoring is labour intensive and time consuming. Therefore, non-destructive image-based techniques, particularly those utilising 3D structural data, have gained attention. We developed a stereo vision-based system to estimate the height of vertically supported tall plants in protected facilities, given plant height serves as a vital measure of crop growth. Our system uses a mobile platform with a top-angle view of a stereo vision depth camera for data acquisition and machine learning in its core for data analysis. First, we collected weekly RGB and depth (RGBD) streams from plant gutters in three glasshouse compartments with different light treatments. We used part of the RGB data collected to train and validate a deep learning segmentation model to detect plant tops and bases. Detected tops and bases of an image were then mapped to the generated 3D scene using the depth image of the same frame. Thresholds and 3D clustering are used respectively to remove background and eliminate outliers in top and base detection mapped to 3D space. Finally, the height of each plant was calculated using the cluster centres of the tops and bases of the plants. Manually measured heights of ten selected plants per environment were used to validate the height estimations. Similar growing patterns were observed between imaged and manually measured plant heights, which showed strong correlations of 0.87, 0.96, and 0.79 R2 scores, respectively, under unfiltered ambient light, Smart Glass film, and shifted light. These promising results demonstrate the feasibility of our proposed method for a vertically supported capsicum crop in a commercial-scale protected crop facility.