Owing to the expansion of target crops of multi-cropping, problems arise in work accuracy; and improvement in versatility can be done by making each work part electric and tweakable. We aimed to build a more optimal control system by detecting the crop amounts in the cutting section and using them as feedback for the control parameters. We created an extremely low-speed reciprocating electric knife. The cutting load was determined based on the feedback system and the power consumption waveform showing the correlation between the cutting load and the crop information under quasi-static conditions. As a result, four parameters were extracted based on the correlation between power consumption and cutting force and the results are discussed.

We built a prototype measurement jig for installing 3D laser scanners at multiple points in pear orchard; further we developed methods for collecting point cloud data and “tree extraction, ” which enable to identify pear trees in the point cloud. The proposed tree extraction algorithm consists of four steps: Create grid; Remove points that lie on the ground surface; Remove points that lie on trellis wires using reflective intensity threshold; and Remove micro-noise via radius outlier removal. We measured 14 pear trees and successfully integrated 13,843,644 points into point cloud. The result obtained via tree extraction method demonstrate that the removal ratios for ground, wire, and micro-noise components were 100 %, 99.6 %, and 87.6 %, respectively, and the tree extraction accuracy was 94.5 %.

Near-infrared (900-1700 nm) hyperspectral imaging was used to visualize the sugar content distribution in cross-sections of the Japanese pear “Nikkori.” The average absorbance spectrum (approximately 11×11 mm) and sugar content for each block of the fruit cross section were measured using a grid gauge. A sugar content estimation model with high prediction accuracy was successfully developed using partial-least-squares (PLS) regression analysis. The prediction coefficient of determination (R² p) between the predicted and measured sugar content was 0.73, with a root mean square error of prediction (RMSEP) of 0.68. The coefficient of determination between the mean value of the sugar-content distribution mapped by this model and the mean value of the measured sugar-content distribution was 0.91.

This study was conducted to develop a labor-saving individual detection method for broiler (meat chicken) breeding management. After individual detection using an object-detection algorithm based on deep learning technology, we evaluated the accuracy of the object detection model. Using a surveillance camera installed on the breeding facility ceiling, a new object detection model was created and operated continuously for about 1 week with regular re-learning using recorded growth images. Consequently, a method for individual detection was configured. Results show that continuous detection for 5-week-old individuals can be achieved by re-learning for this amount of activity and for a weight gain rate of 20 % or less.