Oil Palm Tree Detection Using ArcGIS Deep Learning Tools on Aerial Imageries Obtained from Unmanned Aerial Vehicle
Authors: Badi Hariadi & Tri Haryo Sagoro
Summary
The increasing difficulty of agricultural expansion due to strict regulations on the opening of new oil palm plantations has led to the need for agricultural intensification programs. One of them is to implement the best management practices to increase production and profit in the plantation business. Targeting the allocation of costs and budgets is one form of the best management practices in an oil palm plantation. Conducting a careful and accurate census of oil palm trees, and then using the data, is one way to support these best management practices.
The evaluation method used was a confusion matrix. The results of detection using deep learning tools in the ArcGIS Pro software with the YOLOv3 algorithm with 1 data class, namely ‘Oil Palm’, with the model name: OPDetectionModel_Res10cm_YOLOv3_v11_Stage1_scorexxxx.dlpk, on RGB aerial photos taken using UAVs, obtained an average prediction accuracy of 94.6%; an average precision of 99.7%; and an average recall of 94.9%.
The deep learning model used in this study achieved very good results, so it is expected to speed up the process of palm oil tree census and also be able to provide palm oil tree calculation data with high accuracy and precision.
Keywords: Census, Oil Palm, Deep Learning, ArcGIS Pro, Confusion Matrix, YOLOv3, Aerial Imagery, Unmanned Aerial Vehicle (UAV)
Introduction
Oil palm is one of the most important agricultural commodities in Indonesia. It is used to produce palm oil, which is a valuable commodity that is widely used by the industrial sector. In 2018, the total area of oil palm plantations in Indonesia was 14.33 million hectares with a production of 42.9 million tons. The increase in area and production in 2018 compared to previous years was caused by an increase in the coverage of oil palm plantation companies. It is estimated that in 2019, the total area of oil palm plantations in Indonesia will increase by 1.88 percent to 14.60 million hectares with an increase in CPO production of 12.92 percent to 48.42 million tons (BPS, 2019).
According to the official data of the Government of Indonesia released by the Ministry of Agriculture based on SK №833/KPTS/SR.020/M/12/2019 on the Determination of the Area of Indonesian Palm Oil in 2019, it mentions a figure of 16,381,959 ha as the total area of total oil palm cover in Indonesia (WWF-Indonesia, 2020).
The Presidential Regulation (Perpres) №44 of 2020 states that the revision of the 2015 Indonesian Sustainable Palm Oil (ISPO) principles and criteria will be published within one month of the issuance of the Perpres. This provides hope that the new standards will be able to better accommodate the principles and criteria of intensification, prevention of new plantation expansion, transparency and legal certainty, and conservation protection principles at the site level (WWF-Indonesia, 2020).
The increasing difficulty of agricultural expansion due to strict regulations for sustainable palm oil plantation development has led to the need for agricultural intensification programs. One of them is to implement best management practices to increase production and profit in the plantation business. Achieving the target in allocating costs and budgets is one of the forms of best management practices in a palm oil plantation.
Intensification programs are implemented to improve crop performance, one of which is good block management. Blocks with good performance are blocks that do not experience a gap between target production and realization, while blocks with poor performance are blocks that experience a gap between target production and realization. One of the problems that causes production gaps is the condition of the plants, including the low number of oil palm trees per hectare (Siahaan & Wijaya, 2020).
One of the reasons why the production potential is not achieved is that the plants in a certain area are not entirely productive. Instead, they may be abnormal, dead, or even empty (no plants). Therefore, the insertion of non-productive trees or empty spots is carried out based on the results of a census of trees. To obtain accurate plant data that is in accordance with the actual conditions in the field, it is necessary to conduct a census of trees carefully (Madusari, Sibatuara, & Purwandi, 2014).
One of the benefits of the results of a tree census is as a basis for calculating fertilizer requirements. In order for the cost of fertilization to be effective and efficient, it is necessary to carry out tree census activities carefully, so that they can produce accurate data. (Madusari, Sibatuara, & Purwandi, 2014).
The cost of fertilization is one of the variables that is incurred for each oil palm tree, where the cost of fertilization can reach 60% of the total maintenance cost (Panggabean, Sihombing, & Salmiah, 2013). In addition to fertilization, the costs of maintenance include row maintenance, pruning, and pest and disease control. Therefore, the actual number of trees within a block is important to know in order to avoid cost waste and ensure that the costs incurred are on target.
Current standards for monitoring the number of oil palm trees include deploying workers to conduct direct calculations in the field or manually counting from aerial photos. Information on the number of oil palm trees is a key factor for the management and supervision of oil palm plantations, to obtain yield prediction figures (Daliman, Abu-Bakar, & Nor Azam, 2016).
In addition to the methods mentioned above, another method that can be used to obtain the actual number of oil palm trees in a block of oil palm plantation is the method of automatically counting oil palm trees from aerial photos taken using UAVs or unmanned aerial vehicles using deep learning tools in the ArcGIS Pro software. This automation can replace manual counting in the field, which takes a long time to implement.
This paper is written with the purpose of providing information about the use of the ArcGIS Pro software, namely to detect oil palm trees in aerial photographs taken using UAVs. The paper will also provide information on the performance of a deep learning model that implements the YOLOv3 architecture. Deep learning is a learning method for data that aims to create data representations (abstractions) in a hierarchical manner using a number of data processing layers (Heryadi & Irwansyah, 2020).
ArcGIS Pro software is a Geographic Information System (GIS) software that has deep learning functionality. Meanwhile, UAV or Unmanned Aerial Vehicle is an unmanned aircraft equipped with a camera to take aerial photos in oil palm plantation areas. ArcGIS Pro software has deep learning tools that can be used to help count oil palm trees from UAV aerial photos easily and quickly. The algorithm used to build the oil palm tree detection model is the YOLOv3 algorithm with 1 data class, namely ‘Oil Palm’. The model used is a self-developed model that has been converted into a package called the model: OPDetectionModel_Res10cm_YOLOv3_v11_Stage1_scorexxxx.dlpk.
The samples used as the subject of this study were distinguished based on the type of camera used during aerial photography, at different times, different canopy size/age of oil palm trees, and different locations. In addition, the model was also tested to detect oil palm trees in hilly and bushy areas.
The evaluation method used was the confusion matrix method, which represents the prediction and actual condition (actual) of the data generated by the Machine Learning algorithm. Based on the confusion matrix, we can determine Accuracy, Precision, Recall, and Specificity (Arthana, 2019).
Materials & Method
The material used was 10 cm RGB aerial photos of oil palm plantations obtained from UAVs using several types of cameras, including the Canon S100 camera, the Sony A5000 camera, the Sony A5100 camera, the Sony RX100M3 camera, the DJI Phantom 3 Pro camera, the DJI Phantom 4 camera, the DJI Phantom 4 Pro camera, and the DJI Mavic 2 Pro camera.
The aerial photos were not taken at the same time and in the same place, but were instead archives of aerial photos from 2015 to 2021 with a sample dataset from immature oil palms to mature oil palms, the condition of oil palm trees that are well-maintained and neglected, and on various types of topography. The excerpts of some of the aerial photos used were 4 Ha (200m x 200m).
The specifications of the computer used to detect oil palm trees were an HP ProBook notebook with the following specifications: Intel i7–7500 CPU; 8GB RAM; and Windows10 Pro 64bit OS; without GPU.
The software used to detect oil palm trees is ArcGIS Pro software version 2.7 with the Deep Learning Framework installed. Instructions for installing the deep learning framework can be found in the following tutorial: https://pro.arcgis.com/en/pro-app/latest/help/analysis/deep-learning/install-deep-learning-frameworks.htm. The deep learning tools used were Detect Objects Using Deep Learning, with the parameter used were padding=56; threshold=0.1; nms_overlap=0.5; batchsize=4; exclude_pad_detections=true; Non-Maximum Suppression=enabled; and Max Overlap Ratio=0.5.
The deep learning algorithm model package used in the ArcGIS Pro software is OPDetectionModel_Res10cm_YOLOv3_v11_Stage1_scorexxxx.dlpk, created by the GIS Department of Wilmar International Plantation. The architecture of the deep learning model used is YOLOv3. Training data was carried out using a sample dataset of 448x448 pixels, with a stride size of 112 pixels with a sample size of ±100,000 samples. Of the ±100,000 samples, 10% were used as a validation dataset and 90% were used as a training dataset.
The evaluation method used to calculate Accuracy and Precision is the confusion matrix method (Narkhede, 2018). Oil Palms that are detected as True Positive; palm oil thickets, shrubs, coconut trees, and others that are detected as oil palms as False Positive; while oil palm trees that are not detected as False Negative. The formula used is:
Accuracy = (TP + TN ) / (TP+FP+FN+TN)
Precission = (TP) / (TP+FP)
Recall = (TP) / (TP+FN)
where accuracy is the ratio of the number of samples classified as true positive (TP) to the total number of samples (Chicco & Jurman, 2020). Precision, on the other hand, is the ratio of true positive predictions (TP) to the total number of samples predicted as positive (TP and FP). Recall or sensitivity, on the other hand, is the proportion of all positive classes that are predicted correctly (Ghoneim, 2019).
Results And Discussions
The following image does not show a color composite, but rather shows circles of palm oil canopy that are colored to represent object classes. The size of the circles uses an approach to the size of the palm oil canopy, the data for which is also obtained from the deep learning process that was run. The number of aerial photo samples used is 46 aerial photos.
In the following image, circles of different colors are used to represent the following objects:
· Red circles (TP/ True Positive) are oil palm trees that are automatically detected by the deep learning model;
· Blue circles (FN/ False Negative) are manual digitizing for oil palm trees that are not yet automatically detected by the deep learning model; and
· Yellow circles (FP/ False Positive) are detection errors (errors can be in the form of shrubs that are detected as oil palm trees, bananas that are detected as oil palm trees, or coconut trees that are detected as oil palm trees, etc.).
The terms TP, FN, and FP are used to evaluate performance using the confusion matrix method.
The planting patterns and ages of oil palm planted as samples are also diverse. This is intended to assess the performance of the deep learning model on various types of oil palm planting characteristics.
Using the confusion matrix formula, the average ratio of the samples used is obtained. The table below shows the performance ratio of each sample according to the figure above.
From the above figures and tables, the results of the trial of detecting oil palm trees using a deep learning model were obtained with the following statistical values:
Taking the example of the average statistical value above, it can be read as follows: (the number 1000 is a hypothetical number to make it easier to describe the value in numbers):
· The average accuracy value is 94.6%, which means that: from 1000 oil palm trees in an aerial photo clip, there are 946 oil palm trees that are correctly detected as oil palm trees, and the remaining 54 are errors in predicting non-oil palm objects that are detected as oil palm and errors in the existence of oil palm trees that are not detected.
· The average precision value is 99.7%, which means that: from 1000 oil palm trees predicted as oil palm trees, only 3 objects are incorrectly detected as oil palm trees.
· The average recall value is 94.9%, which means that: from 1000 oil palm trees in an aerial photo clip, there are 51 oil palm trees that are not detected. Or in other words, the manual digitization process of a total of 1000 oil palm trees is as many as 51 trees (949 are automatically detected and 51 trees are manually digitized).
Conclusions
Based on the confusion matrix evaluation method, the detection results using deep learning tools in the ArcGIS Pro software with the YOLOv3 algorithm with 1 data class, namely ‘Oil Palm’, with the model name: OPDetectionModel_Res10cm_YOLOv3_v11_Stage1_scorexxxx.dlpk, on RGB aerial photographs taken using UAVs, the average prediction accuracy value was 94.6%; the average precision value was 99.7%; and the average sensitivity value was 94.9%. This means that the model is able to correctly identify oil palm trees with a high degree of confidence.
From the above statistical figures, the performance of the deep learning model used is also good enough to be used and run in various planting pattern characteristics and oil palm canopy sizes. The lowest sensitivity (minimum recall) is 84%, meaning that when manual digitization is required, manual digitization is performed no more than 16% of the total number of oil palm trees. The low sensitivity of 84% indicates that the model may miss some oil palm trees, but it is still relatively low. This is likely due to factors such as the quality of the aerial photos, the size and density of the oil palm trees, and the variety of oil palm trees.
By using this deep learning model, it is hoped that it can accelerate the digitization process of oil palm trees from RGB aerial photographs taken using UAVs, and can also provide data calculations of oil palm trees with high accuracy and precision. Overall, the results of this study are promising and suggest that deep learning has the potential to be a valuable tool for oil palm management.
Reference
Arthana, R. (2019, April 5). Mengenal Accuracy, Precision, Recall dan Specificity serta yang diprioritaskan dalam Machine Learning. July 13, 2021, dari Medium: https://rey1024.medium.com/mengenal-accuracy-precission-recall-dan-specificity-serta-yang-diprioritaskan-b79ff4d77de8
BPS. (2019). Statistik Kelapa Sawit Indonesia 2019. (S. S. Perkebunan, Ed.) Jakarta: Badan Pusat Statistik.
Chicco, D., & Jurman, G. (2020). The Advantage of the Matthews Correlation Coefficient (MCC) Over F1-Score and Accuracy in Binary Classification Evaluation. Diakses pada July 13, 2021, dari NCBI: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6941312/
Daliman, S., Abu-Bakar, S., & Nor Azam, S. (2016, April 13–14). Development Of Young Oil Palm Tree Recognition Using Haarbased Rectangular Windows. 8th IGRSM International Conference and Exhibition on Remote Sensing & GIS (IGRSM 2016), 37. Diakses pada July 15, 2021, dari https://iopscience.iop.org/article/10.1088/1755-1315/37/1/012041
Ghoneim, S. (2019, April 2). Accuracy, Recall, Precision, F-Score & Specificity, which to optimize on? Diakses pada July 13, 2021, dari towards data science: https://towardsdatascience.com/accuracy-recall-precision-f-score-specificity-which-to-optimize-on-867d3f11124
Heryadi, Y., & Irwansyah, E. (2020). Deep Learning Dan Aplikasinya Di Bidang Informasi Geospasial. Depok: PT. Artifisia Wahana Informa Teknologi.
Madusari, S., Sibatuara, W., & Purwandi, H. (2014, November 3). Perbandingan Metode Sensus Pokok Tanaman Kelapa Sawit Menggunakan Staplecard dan GPS pada Tanaman Menghasilkan Pertama (Studi kasus di PT Citra Sawit Lestari, Kalimantan Utara). Jurnal Citra Widya Edukasi, 6, 45–52. Diakses pada Juli 15, 2021, dari https://journal.cwe.ac.id/index.php/jurnal_citrawidyaedukasi/article/view/127
Narkhede, S. (2018, May 9). Understanding Confusion Matrix. Diakses pada July 13, 2021, dari towards data science: https://towardsdatascience.com/understanding-confusion-matrix-a9ad42dcfd62
Panggabean, R. M., Sihombing, L., & Salmiah. (2013, October). Analisis Pengaruh Biaya Pemeliharaan terhadap Pendapatan Agribisnis Kelapa Sawit (Kasus: Desa Pangkatan, Kecamatan Pangkatan, Kabupaten Labuhan Batu). Neliti, -. Diakses pada July 8, 2021, dari https://www.neliti.com/publications/15133/analisis-pengaruh-biaya-pemeliharaan-terhadap-pendapatan-agribisnis-kelapa-sawit
Siahaan, M., & Wijaya, H. (2020, Juni). Strategi Peningkatan Produktivitas Tanaman Kelapa Sawit Melalui Pendekatan Manajemen Blok Di Perkebunan Kelapa Sawit Skala Luas. Jurnal Agro Estate, 4. doi:https://doi.org/10.47199/jae.v4i1.117
WWF-Indonesia. (2020, April 16). Pernyataan Resmi Posisi Wwf-Indonesia Terhadap Perpres Nomor 44 Tahun 2020. Diakses pada 07 15, 2021, dari WWF: https://www.wwf.id/publikasi/pernyataan-resmi-posisi-wwf-indonesia-terhadap-perpres-nomor-44-tahun-2020
