Figure 1 - uploaded by Alberto Sassu
Content may be subject to copyright.
The experimental field located in Usini, North-Eastern Sardinia, Italy (WGS84-UTM 32N, EPSG 32632 projected coordinate system).
Source publication
The work aims at discovering the potential and the efficiency of Unmanned Aerial Systems (UAS) and Machine Learning (ML) in agriculture scenario, focusing on crop management and agrochemicals distribution optimization in orchard and horticultural cropping systems. The dissertation includes a general introduction, three experimental chapters and a g...
Contexts in source publication
Context 2
... a method applied on images acquired in a vineyard under specific conditions may not work as well in another vineyard or may not even work in the same vineyard as some of those conditions change. In this section, is fully described the proposed methodology summarized in Figure 1. ...
Context 3
... a matter of fact, if the training loss (evaluated on the train dataset) indicates how well the system is learning to perform the object detection on the training set (that is the already known data), the validation loss (evaluated on the validation dataset) explains how much the system can generalize the detection capability on never seen data. Figure 10 shows the training and validation loss values obtained by our system. The number of epochs, which is the number of times the learning algorithm update the model by analyzing the entire training dataset, is used as a temporal scale. ...
Context 4
... detailed results are shown in the second column of Table 4 As shown in Figure 11, while in set 3 the bunches images are usually well defined and easy to detect, in set 4 and, even more, in set 5 there is a greater overlap between different bunches. Furthermore, the prevalence of red grapes in set 3 makes the detection much easier compared to the detection of white grapes, more similar in color to the surrounding vegetation. ...
Context 5
... stated before, the system is not error-free, since some bounding boxes are not detected at all, and others are not correctly detected, meaning that their IoU, with the ground truth, is lower than 0.5. In Figure 12, an example of correct detection on a test image is shown (the green boxes represent the ground truth, while the blue ones are the detection results) since the IoU is clearly greater than 0.5. Other examples (Figure 13), ...
Context 6
... Figure 12, an example of correct detection on a test image is shown (the green boxes represent the ground truth, while the blue ones are the detection results) since the IoU is clearly greater than 0.5. Other examples (Figure 13), ...
Context 7
... some of the typical problems of object detection. In Figure 13a, only one out of two bounding boxes is correctly detected. In Figure 13b and 13c, the two bunches are detected but as a single element. ...
Context 8
... Figure 13a, only one out of two bounding boxes is correctly detected. In Figure 13b and 13c, the two bunches are detected but as a single element. This is one of the cases in which the error can be considered as "less severe", since the area containing the bunches has been correctly detected. ...
Context 9
... when many bunches stay so close together inside the same image, they are difficult to distinguish. In Figure 13d, the picture is out of focus and only the larger of the two bunches has been correctly detected. These examples confirm the results presented in Table 5 since a single bunch is almost always correctly detected. ...
Context 10
... in the previous cases, most of the errors are due to the incorrect detection of overlapping bunches (Figure 14a), others are caused by the inability to correctly detect shaded parts (upper-left box in Figure 14b). Another error is the incorrect detection of some leaves as bunches (Figure 14c) and it is probably due to the difference between the Sardinian grape varieties and those in the training dataset. ...
Context 11
... in the previous cases, most of the errors are due to the incorrect detection of overlapping bunches (Figure 14a), others are caused by the inability to correctly detect shaded parts (upper-left box in Figure 14b). Another error is the incorrect detection of some leaves as bunches (Figure 14c) and it is probably due to the difference between the Sardinian grape varieties and those in the training dataset. ...
Context 12
... in the previous cases, most of the errors are due to the incorrect detection of overlapping bunches (Figure 14a), others are caused by the inability to correctly detect shaded parts (upper-left box in Figure 14b). Another error is the incorrect detection of some leaves as bunches (Figure 14c) and it is probably due to the difference between the Sardinian grape varieties and those in the training dataset. In this case, the difference between the leaves in the image and all of those previously shown, belonging to the GrapeCS-ML dataset, is evident. ...
Context 13
... type of error can be significantly reduced by training and testing the system with images collected in the same geographic area or, even better, in the same vineyard, but the focus of this work was, on the contrary, the analysis of a generalized capability of the framework. This capability is shown, for example, in the two images of the internal dataset in Figure 15, where the same bunches are depicted. It is worth noting that, despite the image in Figure 15a is considerably overexposed, all the clusters have been correctly recognized as in Figure 15b. ...
Context 14
... capability is shown, for example, in the two images of the internal dataset in Figure 15, where the same bunches are depicted. It is worth noting that, despite the image in Figure 15a is considerably overexposed, all the clusters have been correctly recognized as in Figure 15b. The ability to correctly detect grape bunches of varieties never seen before under uncontrolled lighting conditions is the main novelty of this work. ...
Context 15
... capability is shown, for example, in the two images of the internal dataset in Figure 15, where the same bunches are depicted. It is worth noting that, despite the image in Figure 15a is considerably overexposed, all the clusters have been correctly recognized as in Figure 15b. The ability to correctly detect grape bunches of varieties never seen before under uncontrolled lighting conditions is the main novelty of this work. ...
Context 16
... EPSG 4326) at 125 m above sea level. Figure 1 shows an overall view of the artichoke field object of the study. ...
Context 17
... (b) To address the issue of the uniformity or uneven distribution of the growing rate, it is possible to display a heat-map (as in Figure 10) where the spatial distribution of the growing index value for each tracked plant, is shown (color scale) over the full crop field. Figure 11 shows the temporal evolution of two sample artichoke plants (detected with the YOLOv5 network, the behavior with the FPN being the same) in a simple isolated case (a) and a quite common situation (b) where the growing of nearby plants is quickly reaching a size of mutual interference and partial overlaps. ...
Context 18
... (b) To address the issue of the uniformity or uneven distribution of the growing rate, it is possible to display a heat-map (as in Figure 10) where the spatial distribution of the growing index value for each tracked plant, is shown (color scale) over the full crop field. Figure 11 shows the temporal evolution of two sample artichoke plants (detected with the YOLOv5 network, the behavior with the FPN being the same) in a simple isolated case (a) and a quite common situation (b) where the growing of nearby plants is quickly reaching a size of mutual interference and partial overlaps. Another result reported in Table 8, highlight the occupancy rate of the plants along the rows and its development through the growing season. ...
Context 19
... general, the detection rate (recall) is higher in the early period of the crop when the plants are smaller and isolated and is lower in the late period of the year. This behavior highlights the increasing difficulties to detect and distinguish plants in the last phases due to the mutual overlap of the bigger plants within the rows (Figure 10). A similar trend is also visible for the measure of precision which is over 90% in the early dates. ...
Context 20
... a full history of the vegetation process is obtained, for each individual plant (Figure 11), at the different stages of the development process. This detection system's ability will open new scenarios for plant detection, easily allowing the operator to monitor the entire field and evaluate the condition of each plant, specifically for those that show different conditions respect to the rest of the field. ...
Similar publications
Input optimization is a distinguishing characteristic of Precision Agriculture approaches, helping reduce the environmental impact and costs and increase vegetable production quality. Thanks to the high automation evolution of Unmanned Aerial Systems (UAS), a new approach derived from their combination with Deep Learning techniques is leading to si...
