U-Net with Spatial Pyramid Pooling Module for Segmenting Oil Palm Plantations

Palm oil is one of the most important commodities for Malaysia's economy. As the second-largest exporter of palm oil in the world, the government has set up various rules and regulations to promote sustainable plantations. Yet, some parties will take advantage of the rules by expanding their plantation areas beyond the permitted size. Thus, a remote sensing approach to automatically monitor the plantation size is proposed in this paper by using a deep neural network segmentation method. The spatial pyramid pooling (SPP) module is integrated with the well known U-Net architecture to improve the segmentation accuracy. Several variants of U-Net with SPP module are explored through varying the kernel size used in downsampling the input layer. The SPP module is placed right before the bottleneck block between the encoder and decoder sides of the network. The results show that the best accuracy is obtained by using U-Net with SPP of kernel sizes 2, 7 and 14. The proposed method has increased the accuracy from 0.7641 to 0.8152 when tested on Kaggle WiDS Dataset. The increment in performance is attributed to SPP ability in handling various scales input, which is a normal occurrence when the tested images cover a wide range of plantation ages that include young to mature trees.