- Gené Mola, Jordi
- Gregorio López, Eduard
- Sanz Cortiella, Ricardo
- Escolà i Agustí, Alexandre
- Rosell Polo, Joan Ramon
Yield prediction provides valuable information to plan the harvest campaign, fruit storage and sales. Traditionally, yield estimation has been carried out by manual counting of randomly selected samples, without addressing spatial variability within the orchard. To obtain a precise estimation it is necessary to sample a relatively large number of trees, which is unfeasible with manual counting. To solve this issue, this work proposes the use of a Mobile Terrestrial Laser Scanner (MTLS) for fruit detection and yield prediction.
Experimental test were carried out in a commercial Fuji apple orchard. The row of threes was scanned from the two sides (east and west). The measurement equipment consisted of an MTLS comprised of a LiDAR sensor, and a real-time kinematics global navigation satellite system (RTK-GNSS) connected to a rugged laptop. The LiDAR sensor used was a Puck VLP-16 (Velodyne LIDAR Inc., San José, CA, USA), which provides a 3D point cloud with calibrated reflectance values of the measured scene.
The fruit detection algorithm implemented in this work is divided into four steps: (1) Reflectance thresholding, which delete those points presenting a reflectance lower than 60%; (2) Connected Points Clustering using DBSCAN; (3) Fruit separation, which uses a support vector machine (SVM) to predict the number of fruits that contains each cluster; (4) False positive removal, also based on a trained SVM.
From detections obtained with this algorithm, the yield was predicted using a linear model (obtained with training data) that relates the number of detections and the actual number of fruits Three different trials were evaluated: east (E) side scanning, west (W) side scanning and merging data from both scanned sides (E+W). As it was expected, fruit detection results showed lower detection rates when only scanning from one tree side, presenting
detection rates of 38.3% and 48.5% for east and west sides, respectively. However, the detection rate increased up to 75.8% when using E+W data. Similarly, yield prediction results showed higher errors when using data from only one tree side, obtaining a RMSE of 15.2% and 15.3% (east and west, respectively). The prediction improved significantly when using data from both tree sides (E+W), presenting a RMSE of 5.4%.
From these results it is concluded that MTLS has potential in yield prediction in fruit orchards. Although fruit detection rates are moderately successful, the system was able to predict the actual number of fruits with low estimation errors. Only using data from one tree side increases the prediction error, but it has de advantage of reducing a 50% the scanning time, which may be interesting depending on the application and the interest of the farmer. Future works will extend this study to other fruit varieties., This work was partly funded by the Secretaria d’Universitats i Recerca del Departament d’Empresa i Coneixement de la Generalitat de Catalunya and the Spanish Ministry of Economy and Competitiveness under Grants 2017 SGR 646 and AGL2013-48297-C2-2-R. The Spanish Ministry of Education is thanked for Mr. J. Gené’s pre-doctoral fellowships (FPU15/03355). The work of Jordi Llorens was supported by Spanish Ministry of Economy, Industry and Competitiveness through a postdoctoral position names Juan de la Cierva Incorporació (JDCI-2016-29464_N18003).