1. Introduction
Bud Rot (BR) continues to be the primary phytosanitary problem threatening the development of oil palm cultivation in Colombia. This disease has destroyed thousands of hectares of oil palm since it was first reported in the Coldesa plantation (Turbo, Antioquia) in the early 1960s [1]. Since then, the disease has affected more than 160,000 hectares across all production areas of the country [2]. According to Mosquera et al. [3], one way to gauge the potential impact of BR on oil palm cultivation is by demonstrating its effect on fruit production. In Colombia, examples of such effects are numerous. However, one of the most documented cases, due to the significant economic impact it caused, occurred in the southwestern zone. This region’s crude palm oil (CPO) production dropped from approximately 88,200 tons in 2006 to only 8300 tons in 2011 [4], a decrease of 90.5% in 5 years due to the epidemic outbreak of the disease. Its effects were so devastating that they initiated a process of converting susceptible Elaeis guineensis cultivars to inter-specific hybrids O×G (E. oleifera × E. guineensis) resistant to BR. However, it took approximately 15 years for the crude palm oil per hectare rates to match the figures from 2006 again.
Oil palm BR is a disease caused by Phytophthora palmivora (E. J. Butler) [5], a microorganism capable of degrading the internal tissues of the bud in oil palms of any age. Once the infectious process begins, lesions produced by the pathogen become visible 3 to 5 days later as small brown depressions surrounded by a watery advancing zone. Secondary infections are characterized by a more significant number of lesions, eventually affecting more than 80% of the spear over time. With the collapse and destruction of the spear, no new young leaves develop. The palm may die due to the subsequent invasion of saprophytic microorganisms and polyphagous insects, such as Rhynchophorus palmarum, attracted by the strong odor emitted by the decomposing tissue [6,7,8,9].
Although in Colombia, it has been possible to identify some highly productive inter-specific O×G hybrids resistant to BR, such as the Coari × La Mé crossings [10,11], their adoption has not experienced the growth that was expected. According to figures from Sispa [12], the planted area in O×G hybrid cultivars in Colombia was 67,919 hectares, corresponding to only 12.1% of the total oil palm established area. This is a meager figure, considering that 14 years have passed since the first release of this type of cultivar [13]. As a result, ensuring the health and, thus, the sustainability of oil palm cultivation in Colombia remains a considerable challenge due to the number of cultivars sensitive to BR with varying degrees of susceptibility currently present in the country.
Given this scenario, implementing an integrated disease management program by oil palm growers becomes the best possible strategy due to its importance in recovering palms affected by BR and its regional impact in reducing the amount of inoculum available for dispersal. This type of integrated management of BR, which has been technically and economically validated [14,15], consists of two fundamental pillars: preventive and curative. The characteristic of preventive management is that it involves all possible agronomic practices to improve crop conditions without disease. In contrast, curative management consists of a series of practices in the presence of the disease. These practices include early detection of affected palms (censuses), removal of diseased tissue, charring removed tissues, applying preventive and curative chemicals to surrounding palms, and periodic monitoring of intervened palms [16]. However, despite advances in interventions for palms affected by BR, the stage that defines the success of curative management remains early detection of the disease.
Early detection of BR is a practice in which a previously trained operator makes a visual assessment of the oil palm’s spear leaf from the ground to identify initial disease symptoms [17]. This stage is considered crucial because it allows for the rapid intervention of affected palms in the initial stages, which is essential not only to accelerate the recovery process but also to break the chain of infective cycles that occur within the palm bud, and whose effects are only noticeable after the removal of internal tissues. The importance of this stage is such that the most significant variations in BR management costs are attributed to censuses, whose participation can reach up to 91.3% in some cases because the detection of early disease stages requires more frequent censuses [3,18]. However, early detection of BR remains a subjective, unreliable practice highly dependent on the expertise of one or more individuals, given the skills and capabilities of those recognizing the first symptoms of the disease. This increases the uncertainty of curative disease management if the person in charge is not adequately trained to perform the task.
In agriculture, new approaches must be incorporated into traditional disease monitoring and assessment systems to achieve reliable automated detection or diagnosis and streamline disease detection processes with high efficacy. Optical sensors have become an essential device for disease detection in plants due to their ability to capture visual details, their non-invasive nature, the possibility of automation, efficiency in analysis, and the ability to monitor on a large scale at a meager cost [19]. Convolutional Neural Networks (CNNs) are particularly effective for this task due to their ability to accurately classify plant disease after preprocessing the dataset to enhance image quality, remove noise, and normalize the data [20]. For instance, models like DenseNet-121 can automatically extract and learn relevant features from leaf images, improving accuracy through techniques like transfer learning, which allows the system to recognize more vegetables than those in the training dataset [21]. YOLO models, such as YOLOv5, YOLOv7, and YOLOv8, have also been employed for accurate object detection, with YOLOv8 achieving a high mean Average Precision (mAP) score, demonstrating its effectiveness in detecting multiple instances of diseases in an image [22].
Pooling models combined with CNNs can further enhance the efficiency of disease detection by reducing the dimensionality of feature maps, making the models more effective in real-world applications. Transfer learning with models like VGG16, ResNet50, and InceptionNet V3 has been employed to enhance model performance, with ResNet50 emerging as a leading performer in some evaluations [23]. Finally, the use of pre-trained models like EfficientNetV2S, fine-tuned for higher accuracy in detecting plant diseases in challenging conditions, underscores the importance of dataset optimization in deep learning applications for plant disease detection [24].
The use of CNNs, particularly ResNet50, has been highlighted for its high accuracy in plant disease classification, achieving a 90.0% accuracy rate in various diseases [25], as well as in specific pathosystems such as Tomato Bacterial Spot (Xanthomonas spp.), Corn Common Rust (Puccinia sorghi), and Potato Early Blight (Alternaria solani) with 95.5%, 96.7% and 97.6% accuracy rates, respectively [26]. However, despite technological advancements that have improved curative management processes for many economically important crops, success still relies on adequate and objective early disease detection. For BR in oil palm, progress has been minimal, nearly negligible. These tasks remain reliant on the prior knowledge and expertise of individuals performing the task, with a high degree of subjectivity that hinders timely intervention for detected palms. This research aimed to develop an integrated system for digital field monitoring and image analysis to enhance the early detection of Bud Rot in oil palm, enabling more accurate and timely identification of the disease. Implementing these devices seeks to provide technological support to increase the effectiveness and efficiency of curative management in oil palm and thereby enhance the economic sustainability of the crop.
2. Materials and Methods
2.1. RGB Image Acquisition Device
The RGB image acquisition device was based on meeting a series of basic requirements that would meet the needs of working in outdoor environments without limiting the possibilities of obtaining and storing high-quality images at a low cost. As a result, the device consisted of three main components. 1. Body: This component comprised a fully extended 12 m telescopic dielectric fiberglass pole, which allows for adjusting the working height by deploying each of its sections (8 bodies of 1.5 m each) (Figure 1a). 2. Optical sensor for RGB image acquisition: The optical sensor component consisted of a digital camera (HERO 7, GoPro Inc., San Mateo, CA, USA) specially selected for its dimensions and weight (62.3 × 44.9 × 33 mm and 116 g). This camera has the following technical characteristics: (i) sensor type—CMOS (1/2.3 inch); (ii) shutter—rolling; (iii) sensor dimensions—6.17 × 4.63 mm (0.243 × 0.182 in); (iv) focal length—linear FOV 24 mm (min.)–49 mm (max.); (v) image resolution—4096 × 3072; (vi) output type—H.264 MP4 (Figure 1b). 3. Receiver: The receiver component comprised an electronic tablet with a 10.61” screen (M10 Plus, Lenovo Group Ltd., Morrisville, NC, USA), providing better visualization of the target due to the quality of the captured 2K image (2000 × 1200 TDDI LCD) (Figure 1c). The digital camera and electronic tablet did not require external cables or additional accessories for communication, as their connection could be established via Bluetooth.
2.2. Location of the Monitoring Plot and Evaluation of Detection Methods
To evaluate the effectiveness of early detection of palms affected by BR using the proposed device in the field, a monitoring plan was structured on a commercial oil palm plot belonging to GREMCA S.A. (El Copey, Cesar, Colombia), one of the advancing fronts of the disease located in the northern region of Colombia (Figure 2a). The selected plot had a population of 672 African-origin palms susceptible to the disease (E. guineensis) of the Deli × La Mé cultivar, established for 15 years with an average height of 10 m above ground level at the time of evaluation (Figure 2b).
The monitoring plan consisted of monthly evaluations from January to December 2022, where the operator, using the detection device previously described, conducted a palm-by-palm survey (census type) to capture RGB images of the middle and lower third of the spear leaf. Subsequently, after the field acquisition session, images were downloaded onto a laptop and stored with a unique code corresponding to the spatial location of the assessed palm. In parallel with this monitoring and on the same evaluated palms, a second assessment was conducted by plantation personnel in charge of the task through the identification of visible disease symptoms at ground level without using the device and storing the information in digital applications with geographic coordinates. To avoid biases in the classification of the presence or absence of BR in the evaluated palms, the recording of information from both activities was never carried out simultaneously. Finally, all the images obtained were stored in individual folders corresponding to the month of acquisition and shared in a virtual repository for analysis using the following detection methods.
2.2.1. Computer-Assisted Detection
An expert in early detection of the disease classified each of the images (presence or absence) according to the recognition of the visible symptoms of BR and which were previously shared in the virtual repository (Figure 3). Then, the information on the spatial location was stored in a spreadsheet according to the evaluation date, keeping the same unique code previously assigned.
All monthly monitoring records, both those using the RGB image acquisition and analysis device and those made by plantation personnel, were systematized in electronic templates and analyzed in R studio libraries [27]. For the temporal analysis, graphs showing the number of detected cases, monthly incidence curves, and accumulated incidence curves according to each evaluated detection methodology were generated. Additionally, monthly maps were generated for the spatial analysis, showing the distribution of georeferenced points identified as cases detected by each evaluated method. The purpose of both temporal and spatial analysis was to identify delays (deltas) in the time variable and identify, distribute, and locate detections made by the proposed device and those made by plantation personnel.
2.2.2. Automatic Detection Using AI
A deep learning-based model was calibrated to detect the initial lesions of BR in the spear leaf. To accurately outline and identify BR lesions, an instance segmentation model was employed for this computer vision project. The entire modeling process was conducted using Roboflow [28], a web-based platform that offers a suite of tools to streamline the development of computer vision workflows for real-world applications. Following image acquisition and classification of RGB images with initial symptoms of the disease, a dataset comprising 533 images was selected for training the instance segmentation model. BR lesions were annotated by the expert in early detection of the disease using Roboflow’s Smart Polygon tool to enhance the accuracy and efficiency of the labeling process, resulting in the generation of 2389 well-defined labels.
The annotated image dataset was then divided into training, validation, and test sets with a ratio of 76:12:12. To address the need to increase the data available and improve the model, a technique called data augmentation was employed. This technique involves increasing the number of images and labels from those established in the project. It is an effective way to enhance model performance by creating augmented images and adding them to the dataset, which helps the model learn to better identify classes, particularly in conditions that are underrepresented in the dataset. The augmentations applied to the dataset included rotation (between −5° and +5°), saturation (between −25% and +25%), brightness (between −20% and +20%), and exposure (between −10% and +10%) with five outputs per training example. The augmented images were included only in the training set, not in the validation or test sets. This approach ensures that the model is tested and validated on real-world, unmodified data, providing an accurate assessment of its performance. The augmented dataset included 2165 images labeled with BR lesions, resulting in a total of 10,949 labeled instances. The Roboflow 3.0 Instance Segmentation model, with the COCOs-seg checkpoint, was trained on this dataset using the accurate training option. Although Roboflow offers ease of use and remote training, it has not publicly disclosed the technical specifications of this segmentation model, which limits control over its details. Nevertheless, this model was deemed adequate to demonstrate the feasibility of using a deep-learning approach for the early detection of BR lesions (Figure 3).
To evaluate the performance of the model, we utilized three key metrics: precision, recall and mean Average Precision (mAP). Precision is a metric that indicates the accuracy of the model’s positive predictions. It is defined as the ratio of true positive detections to the sum of the true positive and false positive detections. Recall measures the model’s ability to identify all relevant instances of the target class. It is defined as the ratio of true positive detections to the sum of true positive and false negatives. Finally, mAP is a comprehensive metric that combines both precision and recall into a single score. It is calculated by first determining the Average Precision (AP) for each class and then averaging these AP values. For all three metrics, higher values are indicative of better model performance.
2.3. Performance of Detection
The performance (efficiency) of the time spent on the detection processes with and without the use of the device was recorded every month. These values were calculated based on the number of palms visited in an 8 h period, equivalent to a wage, and then converted into ha/wage, using the planting density of E. guineensis as reference, which is 143 palms/ha. Finally, the differences in terms of performance (ha/wage) with and without the use of the device were estimated.
3. Results
3.1. Evaluation of Detection Methods: Automatic Detection Using AI
The results of the training process are summarized through the loss components depicted in Figure 4. The Box Loss graph displays a general upward trend after an initial sharp decline, starting around 3.0, dropping to about 2.5, and then gradually increasing to about 2.75 by the end of training. This pattern suggests that while the model initially improves in localizing the lesions, it may be struggling to refine its outline predictions as training progresses.
The Class Loss plot shows a dramatic drop from about 70 near 0 within the first few epochs, after which it remains consistently low. This indicates that the model quickly learns to classify lesions correctly and maintains this performance throughout training. The Object Loss graph exhibits more volatility, with an initial decrease followed by fluctuations between approximately 1.10 and 1.18. This suggests that the model’s ability to detect lesions improves overall but experiences some instability, possibly due to challenging examples or variations in the training data. The slight upward trend in Object Loss towards the end indicates a need for further optimization. Regarding the overall metrics performance, the mAP was 11.6%, a relatively low value, indicating that the overall performance of the model in terms of both detection and segmentation accuracy is modest. This suggests that the model struggles with accurately identifying and localizing the BR lesions. A higher mAP would typically reflect better model performance, with more accurate and consistent detections across various classes.
The precision score was 29.2%, indicating that when the model makes a positive detection, it is correct approximately 29.2% of the time. While this value is higher than the mAP, it still suggests that a significant portion of the model’s positive detections are incorrect (i.e., false positives). Improving precision would involve reducing the number of false positives and ensuring that the model’s detections are more reliable. Finally, the recall score was quite low (17.1%), reflecting that the model identified only a small fraction of the actual positive instances. This indicates that the model may be missing many true positives, which suggests issues with sensitivity and the ability to capture all relevant instances. Increasing recall would involve enhancing the model’s ability to detect more true positives, which could involve refining its training or improving its feature extraction capabilities. The combination of low mAP, precision, and recall suggests that the model is currently under-performing in both detecting and accurately segmenting BR lesions. The low recall, in particular, highlights that many relevant instances are not being identified, while the precision indicates that when the model does make a detection, it is not consistently accurate.
Figure 5 illustrates the detected instances of BR lesions in the spear leaf. The zoomed-in segmented inferences next to each original image indicate the size and shape of the lesions that the model aims to identify, with varying confidence percentages for each inference. These confidence levels provide insight into the model’s certainty in detecting and outlining the lesions, highlighting areas where the model performs well and areas needing further improvement.
3.2. Evaluation of Detection Methods: Computer-Assisted Detection
3.2.1. Temporal Tracking of Detections with and Without the Use of the Device
Overall, during the 12-month monitoring period, 374 palms (55.7%) of the total evaluated were detected as symptomatic regardless of the method used for detection. The RGB image acquisition and analysis device detected 217 palms (58.0% of the total). In contrast, 157 palms (42.0% of the total detected) were detected with traditional methods, showing a difference of 16.0% between the two detection methods evaluated. However, these differences varied over time depending on the assessment’s month. Regarding disease detection without using the proposed device, the maximum peak of detection was recorded in October with 37 cases detected, while with the RGB image acquisition and analysis device, the maximum peak of detection was also in October but with 64 cases detected, representing a 42.1% difference between both methodologies in the same evaluation month (Figure 6a). However, October was not the only month with such high differences between the two methodologies, as differences of up to 78.6% in detected cases were recorded, which was more evident during peak disease periods in April–May and August–September–October (Figure 6a,b). These differences in the number of detected cases with or without using the RGB image acquisition and analysis device also impacted the disease incidence, especially during the peak disease periods of April–May and August–September–October (Figure 6c,d). When graphing the cumulative incidence curves, the disparities become more noticeable, with values that even exceed 30% at the end of the 12-month evaluation period (Figure 6e,f).
None of the detections made with or without using the RGB image acquisition and analysis device coincided in the same month. When these detections coincided with the same detected palm, the most frequent detection delta was two months, corresponding to 36.9% of the detected cases. This was followed, in order, by detection deltas of 3, 1, 4, 5, 6, and up to 7 months, with percentages of 23.6%, 21.0%, 7.6%, 6.4%, 2.5%, and 1.9%, respectively. Finally, none of the obtained deltas were equal to or greater than eight months.
3.2.2. Spatial Distribution of Detections with and Without the Use of the Device
Figure 6 compares the spatial distribution scenarios of BR between traditional detection by trained plantation personnel and detection with the use of the RGB image acquisition and analysis device in the last month of evaluation (December 2022) when the differences between both detection methods evaluated exceeded 30% of the cumulative incidence of the disease. In Figure 7a, it is possible to visualize the disease’s progression, showing high-intensity spreading from the southern end to the northern end of the evaluated plot. This visualization was not possible when the detections were carried out without using the device (Figure 7b).
3.3. Performance of Detection
The average detection performance without using the device was about 12.3 ha/wage, varying from 15.8 ha/wage in the first month to 8.3 ha/wage in the tenth month. In contrast, the average detection yield with the device was around 2.7 ha/wage, ranging from 1.2 ha/day in the first month to 4.7 ha/day in the twelfth month (Figure 8a). The differences in detection performance decreased over time but were never similar, particularly in months where the work was likely to take longer (Figure 8b).
4. Discussion
In this study, a deep learning-based instance segmentation approach is explored for the early detection of BR lesions in the spear leaf of oil palm trees. Overall, the results indicate that while the model shows potential, substantial improvements are needed to enhance its performance across all metrics. Increasing the quantity and diversity of the training dataset, including more examples and annotations, is crucial to help the model generalize better. However, this increase in the quantity and diversity of the dataset involves several exogenous factors such as the number of palms affected by the disease at a given time (an undesirable condition for commercial plantations that implement all possible measures to reduce the incidence of the disease), the opportunity to acquire the images at early stages of the disease, and the availability of the operator with experience in early detection of BR, among other factors.
Currently, technological advancements in the early detection processes of BR in oil palm are minimal. Only a few reports exist utilizing artificial vision devices [29], geospatial technologies [30], multispectral images [31], and hyperspectral images [32]. All these systems have in common that the acquisition and analysis of images are carried out through orbiting satellites or unmanned aerial vehicles (UAVs) to generate orthomosaics (top-down views of the surface) that serve as input to obtain various vegetation indices that are then related to the phytosanitary status of the palms in the field. However, these technologies lack accuracy and precision because the development of initial disease lesions, barely identifiable to the human eye, occurs localized on the lateral side of spear leaves (immature leaf without opening of leaflets) [6]. For this reason, the analysis of images from orthomosaics is erroneous and misleading if the objective is early disease detection. As the secondary cycles of the pathogen progress inside the spear leaf, the lesions become more profound. It is expected to observe fully extended mature leaves with symptoms of chlorosis and necrosis at the leaf apex, eventually leading to total crown rot of the palms, known as a “crater” palm state [9]. Under this scenario, detecting affected palms using these methodologies is possible; however, we emphasize that these are late disease stages.
Our methodology for acquiring and analysis RGB images was primarily based on using a simple device operated from the ground with the capacity to fully visualize the lateral side of spear leaves of all the palms included in the evaluation. It was demonstrated that this device improves the early detection of BR by identifying a higher number of monthly cases, especially during peak months of the disease, compared to detection methods without using the device. Another comparative advantage of these methodologies was the ability to detect symptomatic palms, mainly with a 2-month lead time before being detected by plantation-trained personnel. These results demonstrate a comparative advantage with a high impact on the recovery efforts of detected palms, considering that P. palmivora generates secondary cycles between 36 and 48 h once the infectious process begins [33], making it one of the most feasible options for early detection in the curative management of the disease.
On the other hand, this type of research helped determine the underestimation percentages of the disease when detections are based on identifying visible symptoms at ground level, which, according to the monthly monitoring conducted, is much higher during peak disease months (April–May and August–September–October), with differences of up to 78.6% of undetected cases. These results demonstrate that a curative approach to the disease has been employed for a long time with a high degree of subjectivity, leading to a gradual increase in reinfections in areas where the effects of curative management should be immediate. Ultimately, the palm oil grower abandons disease management plans due to the relentless rise in undetected active foci, resulting in wholly overwhelmed disease management costs. This same phenomenon was seen in the development of epidemics in Tumaco and Puerto Wilches between 2006 and 2009 [34].
In terms of performance, the estimated time for the personnel in charge of detecting BR without using the device falls within the acceptable performance range (7–13 ha/wage) for carrying out the work [18]. However, in our study, we showed that these values tend to fluctuate depending on the prevalence of the disease in the area being evaluated. This is probably due to more cases being detected during the peak months of the disease (August–September–October), and because each operator requires enough time for in situ confirmation of the observed symptoms. While the RGB image acquisition and analysis device may initially seem to have drawbacks compared to traditional field detection methods, it does not suffer from variations over time that may be related to the level of incidence of the evaluated plot. Our study suggests that any initial variation in the device’s performance is likely associated with the learning curve during the early evaluation stages. After 9 months of training and adaptation in the use of the device, it is expected to reach its maximum performance point and stabilize over time (4.7 ha/wage), irrespective of the degree of affectation of the evaluated plot.
5. Conclusions
This research aimed to develop an integrated system for digital field monitoring and image analysis, testing two detection methods: computer-assisted detection and automatic detection using AI as a way to automate the detection work without the help of qualified personnel in the field. However, in light of the results obtained, it is clear that automatic detection using AI requires strengthening the analysis systems with a considerable amount of images of initial lesions of the disease, many more than those analyzed here, to obtain acceptable results.
Despite these results, the detection method as computer-assisted method, the traditional detection method, proved a significant improvement in the early detection processes of BR in oil palm at the adult stage, especially during the peak months of the disease. The use of this device improves the effectiveness of detection by up to 78.6% during peak months, with time deltas at mostly 2 months of difference compared to traditional detection methods. An even more detailed spatial analysis of the data obtained enabled a comprehensive visualization of the front of the disease’s advance, a condition that would be impossible to visualize if only traditional detection methods were available. However, it is important to note that the performance of this type of device is low, which would imply a higher cost in the curative management of the disease and a low adoption of the technology, for which the incorporation of automatic detection using AI to the proposed device becomes more important will substantially improve field performance. Another option could be a hybrid version of both methodologies where the field operator uses the device only on occasions where there is no certainty about the presence of the initial symptoms of BR at first sight.
Whatever the scenario, the implementation of the RGB image acquisition and analysis device provides important technological support in improving the effectiveness of early detection of the disease. Its use will allow for shorter recovery times due to less in-depth interventions, lower probability of recurrence and secondary infections, greater effectiveness in removing affected tissue, and finally, a decrease in the amount of inoculum available in the environment, making it a highly feasible option for curative management of the disease and the economic sustainability of the oil palm crop. However, the degree of adoption of this type of technology will depend solely and exclusively on the efficiency with which the work is carried out, for which it is expected that automatic detection using AI will play an important role in the future.
Future Work Consideration: This study demonstrates the potential of AI-based models for detecting early symptoms of Bud Rot (BR) in oil palm through instance segmentation. While the results show moderate performance, several opportunities for future work remain to improve the system. One key area is the exploration of state-of-the-art architectures known for superior performance in real-time detection tasks, which could enhance accuracy and efficiency compared to the current model. Future studies could also expand and diversify the dataset to cover a broader range of conditions and lesion appearances. Improving ground truthing protocols and refining model architectures to handle variability in image quality will be crucial for enhancing performance metrics such as precision and recall. Ultimately, these improvements could lead to more reliable and robust tools for early detection of Bud Rot, with potential applications in large-scale disease management in oil palm plantations.
Conceptualization, J.M.L.-V. and C.B.-A.; methodology, J.M.L.-V. and C.B.-A.; software, D.A.G.C.; formal analysis, J.M.L.-V. and C.B.-A.; investigation, J.M.L.-V. and D.A.G.C.; resources, G.A.S. and A.M.-R.; data curation, J.M.L.-V. and D.A.G.C.; writing—original draft preparation, J.M.L.-V., G.A.S. and C.B.-A.; writing—review and editing, A.M.-R.; visualization, J.M.L.-V. and D.A.G.C.; supervision, G.A.S. and A.M.-R.; project administration, G.A.S. and A.M.-R.; funding acquisition, G.A.S. and A.M.-R. All authors have read and agreed to the published version of the manuscript.
The data presented in this study are available on request from the corresponding author.
The authors express their gratitude to the microbiology student Andrea del Pilar Romero Martinez, as well as to the field assistants Eduardo Manuel Pertuz Cantillo and José Alberto Charris Vargas for their contributions to the execution of disease monitoring activities in the evaluation plot. To the technical and administrative staff of GREMCA S.A. for providing their facilities to carry out this research. To the researcher in the Validation area of Cenipalma, Camilo Jose Ardila Badillo, for his support in the analysis of the performance of the detection tests.
The authors declare no conflicts of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.
Footnotes
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Enlarge this image.Figure 1. Components of the RGB image acquisition device. (a) The 12 m telescopic dielectric fiberglass pole; (b) GoPro HERO 7 digital camera; (c) Lenovo M10 Plus electronic tablet.
Enlarge this image.Figure 2. Spatial location of the study site. (a) Departments comprising the northern region of Colombia. ANT: Antioquia; COR: Córdoba; SUC: Sucre; BOL: Bolívar; ATL: Atlántico; MAG: Magdalena; CES: Cesar; GUA: Guajira. The white box shows the geographical location of the GREMCA S.A. plantation in the disease-advancing zone. (b) Distribution of all commercial oil palm plots in the plantation. The white box shows the spatial location of the selected plot based on the highest accumulated incidence of PC recorded at the time of the study.
Enlarge this image.Figure 3. Flowchart illustrating the protocols used for RGB image acquisition, classification, and analysis in computer-assisted detection and automatic detection using AI.
Enlarge this image.Figure 4. Training (a) Box, (b) Class, and (c) Object loss components for the calibrated instance segmentation model over 127 epochs.
Enlarge this image.Figure 5. Sample outputs from the segmentation model for RGB image processing and analysis.
Enlarge this image.Figure 6. Monthly dynamics of BR detections in oil palm at the adult stage with or without using the device for acquisition and analysis of RGB images. (a) Monthly counts of oil palms with BR detections with or without using the device; (b) Monthly difference in the number of oil palms with BR detections with or without using the device; (c) Incidence curve of BR with or without using the device; (d) Difference in percentage of BR with or without using the device; (e) Cumulative incidence curve of BR with or without using the device; (f) Difference in percentage of BR with or without using the device.
Enlarge this image.Figure 7. Comparison of the spatial distribution of BR cases between detection performed with the use of the device (a) and traditional detection performed without the use of the device (b). Red and green dots represent symptomatic or asymptomatic palms depending on the detection method used, respectively.
Enlarge this image.Figure 8. Performance of BR detection with and without using the device (a) and the difference in performance using both methodologies (b).
References
1. Rojas, E.; Ruiz, E. Investigaciones Sobre La Enfermedad Pudriciones de Cogollo-Pudrición de Flecha de La Palma Africana En La Plantación La Arenosa de Coldesa S.A.; Instituto Geográfico Agustín Codazzi: Bogotá, Colombia, 1972.
2. Romero, H.M. En Qué va El Mejoramiento de Cara a Las Problemáticas de La Palmicultura. Palmas; 2022; 43, pp. 102-107.
3. Mosquera, M.; Fontanilla, C.; Zuñiga, E.; Escobar, G.; Cadena, Y.; León, N.; Velasco, C. Una Experiencia de Coordinación de Acciones Para Enfretar La Pudrición Del Cogollo: Costos Asociados a Su Manejo Curativo. Palmas; 2017; 38, pp. 51-62.
4. Fedepalma. Anuario Estadístico 2012. La Agroindustria de La Palma de Aceite En Colombia y En El Mundo; Fedepalma: Bogotá, Colombia, 2013.
5. Butler, E.J. Report of the Imperial Mycologist, 1918–1919; Science Report Institute Pusa: Calcutta, India, 1919; 82p.
6. Martinez, G.; Torres, G. Presencia de La Pudrición de Cogollo de La Palma de Aceite (PC) En Plantas de Vivero. Palmas; 2007; 28, pp. 13-20.
7. Martínez, G. Identificación Temprana y Manejo Integrado de La Enfermedad Pudrición Del Cogollo. Palmas; 2009; 30, pp. 63-77.
8. Sarria, G.; Martínez, G.; Varón, F.; Drenth, A.; Guest, D. Nuevas Evidencias Del Cumplimiento de Los Postulados de Koch En El Estudio de Las Relaciones Phytophthora Palmivora y La Pudrición Del Cogollo de La Palma de Aceite. Palmas; 2013; 34, pp. 73-83.
9. Torres, G.; Sarria, G.; Martínez, G.; Varón, F.; Drenth, A.; Guest, D. Bud Rot Caused by Phytophthora Palmivora: A Destructive Emerging Disease of Oil Palm. Phytopathology; 2016; 106, pp. 320-329. [DOI: https://dx.doi.org/10.1094/PHYTO-09-15-0243-RVW]
10. Romero, H.M. Hacia La Obtención de Materiales Más Productivos, Resistentes a Enfermedades y Adaptados a Las Condiciones Colombianas. Palmas; 2017; 38, pp. 158-165.
11. Avila-Diazgranados, R.; Daza, E.; Navia, E.; Romero, H. Response of Various Oil Palm Materials (Elaeis Guineensis and Elaeis Oleifera × Elaeis Guineensis Interspecific Hybrids) to Bud Rot Disease in the Southwestern Oil Palm-Growing Area of Colombia. Agron. Colomb.; 2016; 34, pp. 74-81. [DOI: https://dx.doi.org/10.15446/agron.colomb.v34n1.53760]
12. Sispa Áreas En Desarrollo y En Producción. Available online: https://sispaplus.fedepalma.org/ (accessed on 29 March 2024).
13. Ayala, I.M.; Romero, H.M. Cultivares Híbridos OxG y La Reactivación Productiva de Zonas Problema Con PC. Proceedings of the XV Reunión Técnica Nacional de Palma de Aceite; Cenipalma: Bucaramanga, Colombia, 2019; 39.
14. Arias, N.; Beltrán, J.A.; Guerrero, J.M.; Sánchez, C. Tecnologías Para El Manejo de La Pudrición Del Cogollo (PC) de La Palma de Aceite Validadas En Las Zonas Palmeras de Colombia. Palmas; 2014; 35, pp. 39-52.
15. Mosquera, M.; Evans, E.; Grogan, K.; Fontanilla, C. Un Modelo de Simulación Discreto Para Determinar La Edad Óptima de Replantación En Presencia de La Pudrición Del Cogollo. Palmas; 2014; 35, pp. 19-32.
16. Martínez, G.; Arias, N.; Sarria, G.; Torres, G.; Varón, F.; Noreña, C.; Salcedo, S.; Aya, H.; Ariza, J.; Aldana, R. et al. Manejo Integrado de La Pudrición Del Cogollo (PC) de La Palma de Aceite; Cenipalma Fedepalma: Bogotá, Colombia, 2009.
17. Sarria, G.; Mestizo, Y.; Padilla, J.; Zúñiga, F.; Castillo, S.; López-Vásquez, J.M.; Vélez, D.; González, D.; Esquivel, L.; Charris, J. et al. Guía de Bolsillo: Monitoreo y Manejo Curativo de La Pudrición Del Cogollo En Elaeis Guineensis e Híbrido Interespecífico O×G En Colombia; Corporación Centro de Investigación en Palma de Aceite, Cenipalma: Bogotá, Colombia, 2024.
18. Fontanilla, C.; Mosquera, M.; Ruíz, E.; Sánchez, A.; Beltrán, J.; Arias, N. Estimación de Costos de Manejo de La Pudrición Del Cogollo (PC) de La Palma de Aceite. Palmas; 2014; 35, pp. 23-37.
19. Mahlein, A.-K. Plant Disease Detection by Imaging Sensors—Parallels and Specific Demands for Precision Agriculture and Plant Phenotyping. Plant Dis.; 2016; 100, pp. 241-251. [DOI: https://dx.doi.org/10.1094/PDIS-03-15-0340-FE] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30694129]
20. Kumari, S.; Pati, N.; Gourisaria, M.K.; Bilgaiyan, S. Application of Advance Deep Learning Models For Plant Disease Detection. Proceedings of the 2024 International Conference on Integrated Circuits and Communication Systems (ICICACS); Raichur, India, 23–24 February 2024; pp. 1-6.
21. Saranya, K.; Nasrin Begun, A.; Archana, A.; Kaviya, N.; Thamarai Selvan, M. Diagnostics System for Plant Leaf Diseases Using Convergence of Computer Vision and Deep Learning. Int. J. Sci. Res. Eng. Manag.; 2024; 8, pp. 1-5. [DOI: https://dx.doi.org/10.55041/IJSREM30313]
22. Ali, U.; Ismail, M.A.; Ariyaluran Habeeb, R.A.; Ali Shah, S.R. Performance Evaluation of YOLO Models in Plant Disease Detection. J. Inform. Web Eng.; 2024; 3, pp. 199-211. [DOI: https://dx.doi.org/10.33093/jiwe.2024.3.2.15]
23. Biswas, T.; Pati, S.K.; Sarangi, S. Plant Disease Detection. Prospects of Science, Technology and Applications; CRC Press: London, UK, 2024; pp. 250-256.
24. Hamed, B.S.; Hussein, M.M.; Mousa, A.M. Plant Disease Detection Using Deep Learning. Int. J. Intell. Syst. Appl.; 2023; 15, pp. 38-50. [DOI: https://dx.doi.org/10.5815/ijisa.2023.06.04]
25. Durga, P.; Gangishetty, D.; Godavarthi, D. Convolutional Neural Network (CNN) Based Identification of Crop Diseases Using ResNet-50. Proceedings of the 2023 International Conference on Sustainable Communication Networks and Application (ICSCNA); Theni, India, 5–17 November 2023; pp. 1293-1301.
26. Awad Bin, N.; Biswaranjan, S.; Abdelhamid, Z.; Maaliw, R.; Islam Sudman, S.; Das, D.; Almeida, F.; Sakr, H. Detecting Three Different Diseases of Plants by Using CNN Model and Image Processing. J. Electr. Syst.; 2024; 20, pp. 1519-1525. [DOI: https://dx.doi.org/10.52783/jes.1455]
27. R Core Team. R: A Language and Environment for Statistical Computing. Available online: https://www.r-project.org/ (accessed on 31 July 2024).
28. Roboflow Everything You Need to Build and Deploy Computer Vision Models. Available online: https://roboflow.com/ (accessed on 30 April 2024).
29. Montero, D.; Rueda, C. Detection of Palm Oil Bud Rot Employing Artificial Vision. IOP Conf. Ser. Mater. Sci. Eng.; 2018; 437, 012004. [DOI: https://dx.doi.org/10.1088/1757-899X/437/1/012004]
30. Viera-Torres, M.; Sinde-González, I.; Gil-Docampo, M.; Bravo-Yandún, V.; Toulkeridis, T. Generating the Baseline in the Early Detection of Bud Rot and Red Ring Disease in Oil Palms by Geospatial Technologies. Remote Sens.; 2020; 12, 3229. [DOI: https://dx.doi.org/10.3390/rs12193229]
31. Alemán-Montes, B.; Henríquez-Henríquez, C.; Largaespada-Zelaya, K.; Ramírez-Rodríguez, T. Evaluación de Flecha Seca En Palma Aceitera (Elaeis Guineensis Jacq.) Mediante Imágenes Multiespectrales, Costa Rica. Agron. Mesoam.; 2022; 33, 47557. [DOI: https://dx.doi.org/10.15517/am.v33i2.47557]
32. Kurihara, J.; Koo, V.-C.; Guey, C.W.; Lee, Y.P.; Abidin, H. Early Detection of Basal Stem Rot Disease in Oil Palm Tree Using Unmanned Aerial Vehicle-Based Hyperspectral Imaging. Remote Sens.; 2022; 14, 799. [DOI: https://dx.doi.org/10.3390/rs14030799]
33. Sarria, G.; Martínez, G.; Varón, F.; Drenth, A.; Guest, D. Histopathological Studies of the Process of Phytophthora Palmivora Infection in Oil Palm. Eur. J. Plant Pathol.; 2016; 145, pp. 39-51. [DOI: https://dx.doi.org/10.1007/s10658-015-0810-9]
34. Silva, Á.; Martinez, G. Plan Nacional de Manejo de La Pudrición Del Cogollo Fedepalma—Cenipalma. Palmas; 2009; 30, pp. 97-121.
© 2024 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.
Details
; García Cárdenas, Diego Alejandro 2
; Bojacá-Aldana, Carlos 2
; Greicy Andrea Sarria 1
; Morales-Rodríguez, Anuar 1
1 Pest and Disease Program, Colombian Oil Palm Research Center, Cenipalma, Bogotá 111121, Colombia;
2 Agronomy Program, Colombian Oil Palm Research Center, Cenipalma, Bogotá 111121, Colombia;