In this article, we employed computer vision and deep learning techniques to select and plant healthy billets, which increased plant population and the yield per hectare of sugarcane planting. We employed well-known convolutional neural network architectures to process large image data sets and transfer learning techniques to expand the results to different sugarcane varieties. It would be very time consuming to collect and label large data sets for each sugarcane variety, for which quality inspection is needed, prior to planting. We used a two-step transfer learning process to extend the trained architecture to new varieties. We compared results obtained during transfer learning using AlexNet, VGG-16, GoogLeNet, and ResNet101 architectures to classical computer vision methods. Our goal was to determine the best approach to detect damaged and good billets in the shortest processing time. Best results in both time and accuracy were obtained with AlexNet. For AlexNet, we compared the permutations of three sugarcane varieties in order to find the best model to identify the healthy sugarcane billets. We then reduced the number of images employed to retrain the model to determine the tradeoff between time and performance. Ultimately, one needs only a few dozen billets of the new variety to retrain the network. Our approach led to meaningful increments in the yield per hectare ranging from 33 to 80% depending on sugarcane variety.
18 July 2018 in IEEE Robotics and Automation Letters
In much of the world, sugarcane is planted in a mechanized fashion using billets, which are shorter segments of cane harvested and cut by a combine harvester. The mechanized harvesting process can damage billets, which introduces pathways for disease, and overall reduction of billet quality. Compared to whole stalk planting with manual methods, growers must approximately double the planting density when using billets. As a first step toward improving sugarcane production with robotics technologies, this letter presents the analysis of sugarcane billet quality using computer vision. A large sample of sugarcane billets was harvested at a research farm in Houma, LA, USA. A group of crop scientists and growers then categorized the billets into six classes of damage according to physical features visually evident. To better understand the correlation between the type of damage and sugarcane germination, we planted 120 samples from each class in test plots and then recorded plant emergence rates. A data set of billet imagery was collected with charge-coupled device and stereovision sensors in both outdoor and indoor lighting conditions. Offline image processing resulted in approximately 90% successful classification of sugarcane billet damage.
Symptoms of red stripe disease caused by Acidovorax avenae subsp. avenae in Louisiana between 1985 and 2010 were limited to the leaf stripe form which caused no apparent yield loss. During 2010, the more severe top rot form was observed, and a study was initiated to investigate the distribution of red stripe in the field and determine its effects on cane and sugar yields. Soil properties data, red stripe incidence and sugarcane yields were all highly variable and were not randomly distributed in the field. Combined harvest data showed a negative correlation between yield components and red stripe incidence with the strongest relationship between sucrose per tonne and disease incidence. Red stripe incidence was positively correlated with several soil properties, including phosphorus, potassium, zinc and calcium. Red stripe incidence also was found to increase with increasing N rate, with the greatest effects in heavy soils. Results also indicated that using red stripe infected cane as a seed source can sig...
Remote sensing techniques and the use of Unmanned Aerial Systems (UAS) have simplified the estimation of yield and plant health in many crops. Family selection in sugarcane breeding programs relies on weighed plots at harvest, which is a labor-intensive process. In this study, we utilized UAS-based remote sensing imagery of plant-cane and first ratoon crops to estimate family yields for a second ratoon crop. Multiple families from the commercial breeding program were planted in a randomized complete block design by family. Standard red, green, and blue imagery was acquired with a commercially available UAS equipped with a Red–Green–Blue (RGB) camera. Color indices using the CIELab color space model were estimated from the imagery for each plot. The cane was mechanically harvested with a sugarcane combine harvester and plot weights were obtained (kg) with a field wagon equipped with load cells. Stepwise regression, correlations, and variance inflation factors were used to identify the best multiple linear regression model to estimate the second ratoon cane yield (kg). A multiple regression model, which included family, and five different color indices produced a significant R2 of 0.88. This indicates that it is possible to make family selection predictions of cane weight without collecting plot weights. The adoption of this technology has the potential to decrease labor requirements and increase breeding efficiency.