Proximal sensing of horticultural crops with mobile device LiDAR

Robust, multipurpose 3D plant model datasets are critical for the development of phenotyping and structural characteristic estimation methods. This study details the construction of a multi-crop 3D point cloud dataset collected in-situ, tests the efficacy of HSV thresholding on colorized point cloud data to segment target plants, and evaluates the efficacy of mobile device LiDAR sensors for plant structural data collection. Across the agricultural season, a mobile device LiDAR sensor was used to collect 295 colorized point clouds of pepper, tomato, and watermelon plants. Segmentation with HSV thresholding and processing resulted in an mIoU score of 89%. Spatial accuracy of the sensor was assessed through comparison of physically measured and sensed tomato plant height, yielding a strong regression fit (R2 = 0.95), but with significant error (RMSE = 5 cm). To assess the sensor’s structural characteristic estimation capabilities, aboveground vegetative biomass, leaf and stem biomass, and leaf and stem number were physically measured in tomato and watermelon plants. Convex hull volume and voxel volume at multiple voxel sizes were related to biomass and structure counts using NLS regression. Regressions between plant volume and structural characteristics showed a strong relationship between each crop and the structural characteristics (R2 = 0.86-0.98). Relative error was high in tomato estimates (RRMSE = 22-38%), but comparable in performance to other sensors in watermelon estimates (RRMSE = 12-23%). The spatial and structural characteristic estimation error was attributed to the sensor and scanning software, restricting mobile device LiDAR’s data collection utility with current technology.