Classification of Oil Palm Growth Status with L Band Microwave Satellite Imagery

With the rapid increase of human population, the demand for foodstuff such as vegetable oil has also risen. Palm oil is one of the vegetable oils with highest yield to land ratio and unsustainable development of plantations is highly discouraged. Therefore, optimization of plantation yield has to be done with good plantation management and microwave remote sensing can be used to provide information to plantation owners to perform suitable actions. However, there is currently a lack of understanding in microwave remote sensing in the context of oil palm and studies need to be done to translate microwave satellite imagery to meaningful information for plantation owners. There are previous studies done to classify microwave satellite imagery with the use of machine learning techniques such as support vector machine (SVM) and convolutional neural network (CNN). In this work, we study the use of machine learning on microwave satellite imagery to classify growth status of oil palms. We first studied the oil palm biophysical parameters together with the use of theoretical model based on radiative transfer (RT) equation where we were able to identify the key parameters and components that are sensitive to L band microwave imagery obtained from ALOS-PALSAR 2. With this knowledge in hand and feedback from agronomists, we then design 5 classification classes that can suitably portray oil palm growth status with respect under L band microwave imagery. We then proceed to classify the image with ground truth data obtained from the field study. With inspiration from techniques such as generative adaptive network (GAN), we generated simulation data with the use of RT model in order to supplement additional data to the CNN and combine this with SVM to establish the hybrid classification model. By getting the best of both techniques, the hybrid classification model managed to achieve 90% training accuracy. The classified images are then overlaid with Google Earth interface and we qualitatively evaluate the classification results with ground study information and census data. We found that the image of the hybrid classification model was in good agreement with ground truth data.