Oil palm is known as one of the important resources in Malaysia and is exported to other countries which can be used to other countries which can be used to produce many palm oil products. To handle this demand, the quality and quantity of oil palm tree plantations are the main issue for the producers, thus proper plantation monitoring like tree detection and counting are required to ensure optimum production. The traditional way to monitor oil palm plantation is by arranging human labour to check every area accordingly but this is known to be time, energy and cot consuming, plus less efficient. Therefore, remote sensing can be utilised to monitor the oil palm plantation effectively, Remote sensing imaging is the action of getting images from a distance. Through satellite imaging, RGB-based image with a detailed, high-resolution information can be obtained, sometimes at no cost, representing a wide coverage area which can be difficult to be reached by humans. However, RGB-based satellite image has its limitations, thus a proper algorithm has to be used to process these images. Hence, this research proposes a method to monitor oil palm tree plantation using Backpropagation Neural Network (BPNN) from RGB-based satellite image. Here, RGB-based satellite image on the selected oil palm plantation area is obtained from a free access platform, Google Earth Pro, and then processed with cropping operation to divide into detailed parts using "Photo" application. All image processing is performed in MATLAB where first, the cropped image undergoes noice removal using median filter as it has better performance to keep all the useful detail in the image. Then, the clean image is converted into grayscale image and region-based segmentation based on the threshold value obtained from the histogram is performed so that it can detect the presence of all oil palm trees from the image. Morphological operation is conducted after that which includes hole filling, opening, closing, and removal of unnecessary object to get better trees detection from the binary image. Watershed transform based on topological principle is also conducted to separate neighbouring trees. In the classification stage, the BPNN is built with the network structure of one input layer, three hidden layers, one output layer, bias of bipolar sigmoid activation function, and learning rate of 0.1. Finally, a validation accuracy is calculated which deals with TP, FN, and FP to compare with the ground truth data. Overall, the final results demonstrate that the accuracy of the tree detection and counting using BPNN algorithm is 95.62%, with precision, recall, and F-measure are 85.94%, 84.05%, and 84.92% respectively. Further testing with other validated oil palm plantation also gives a good accuracy of 82.96%. These findings discover the potential usage of free RGB images from satellite imagery platform to perform oil palm plantation monitoring, specifically tree detection and counting, compared to other commercial satellites, which is very beneficial especially for smallholders.
