Artificial Intelligence Technology Implementation to Identify Palm Trees (Case Area: Indonesia)
SuperMap Software Co., Ltd in collaboration with PT. Independent Research & Advisory Indonesia (IRAI) cooperates in the implementation of Artificial Intelligence (AI) technology to identify palm trees and to count the number of palm trees in plantation area managed by PT. Perkebunan Nusantara V. This project is carried out in 9 main palm plantations owned by PT. Perkebunan Nusantara V in Riau Province, spread across Kampar, Rokan Hilir, Rokan Hulu, and Siak Regencies, covering ± 31.860 hectares.
In the first stage of the project, an identification of palm trees for 5.000 hectares was carried out using SuperMap iDesktopX with an Artificial Intelligence extension. This process was completed in less than 5 working days. PT. Independent Research & Advisory collected orthophoto data with the resolution of 7 cm per pixel as the data input for the following process. The process of identification begins with creating sample data (labels) for each characteristic of palm trees. Palm characteristics in the detection area play an important role in the process of making labels. Two characteristic of palm trees (large and small) in the detection area needs to be identified through the process.
In order to reduce manual workload, the label data format is a box. The box is drawn on the position of each tree. Under the attribute table, the category field is created to define the category of the object.
Create Training Data. Because of the volume of the original image data is large and the convolutional neural network has high requirement for running memory, the image needs to be processed into suitable data for machine learning model training. The purpose of creating training data is to simultaneously cut the image data and label data into small pictures and generate suitable data for neural network model training. Test pictures are generated from left to right, from top to bottom until the test area range is covered. Each small picture is accompanied by a file (*.xml) which records the palm tree location (label) in the picture.
Model Training. The next step of palm trees identification is model training. The purpose of model training is to use the generated training data for neural network model training, then iteratively evaluate the model continuously to obtain usable neural network model. The model is trained by using the generated training data. YOLO model is selected to set up and predict a type of result. The training saves the model once every iteration (epoch) is completed, so that the change of the result will be recognized easily. The model is optimized by loading the pre-training model and adjusting the parameters of machine learning. Model training require longer time since it depends on user’s system configuration, samples, and model training epoch until we find the most suitable model for identification or until the sample’s loss less than 0.1 or much smaller.
Object Detection. The purpose of the object detection tool is to find the target of interest in the image based on the neural network model. On the first stage of the project, SuperMap team use partition image of the whole 5.000 hectares area which covered ± 16 hectares per image from PT. Independent Research and Advisory Indonesia (IRAI). Therefore, object detection is carried out around 5-10 minutes for each partition image. The number count of each identification varies from 300 to 7.000 trees count, depending on the location area of the image.
It marks the targets in the form of rectangular boxes for following spatial statistics and analysis. The dataset can be exported to the necessary format data (*.shp, *.cad or else). The following image is the result of object identification for each image partition.
After the object detection process is complete, PT. Independent Research and Advisory Indonesia (IRAI) conducted Quality Control. They used manual quality control process with a palm consulting team to define the object/error on the detection area. They are also using the Normalized Difference Vegetation Index (NDVI) method by calculating the greenness value of palm. The resulting output will provide an information of the average values per tree. The result will be compared with the result of AI detection which display on the following image
Note: We are looking for distributors, resellers, and partners all over the world. For any inquiry, please contact us at: [email protected]