DenseNet with Spatial Pyramid Pooling for Industrial Oil Palm Plantation Detection

Abstract

Palm oil is the most consumed vegetable oil with many proven benefits to human health. Apart from domestic usage, it has also been widely adopted as a source for biofuel, which produces less carbon footprint compared to the normal fuel. Because of its popularity, a lot of countries has systematically planted oil palm trees to make sure that it will be sustainable for the future generation. Rigorous monitoring of the land used for oil palm plantations is an important step in sustainable farming. Remote sensing approach through satellite imagery has been used to detect the industrial plantations, where their size can be inferred from the segmented regions. Therefore, a good monitoring system relies heavily on accurate detection of the oil palm trees from the satellite images. However, it is hard to detect the plantations because of non-uniformity in the age of oil palm trees, where a young plantation will have sparse canopy pattern, while a mature plantation will have dense canopy pattern. Besides that, any satellite image is prone to the problem of cloud occlusion and thus, raise the difficulty in detecting the plantations. This work presents an improved DenseNet with spatial pyramid pooling (SPP), which can detect the plantations across various scales so that plantation age is not a limiting factor. A set of concatenated features from three different maximum pooling kernels are used to replace the original global average pooling operation. The results show that the addition of SPP module has increased the detection accuracy by 1.78%. Besides that, the results also show that the shallowest DenseNet with 121 layers performs the best compared to the deeper DenseNet versions. The proposed algorithm can be extended to include semantic segmentation module so that the plantation size can be predicted instead of just the presence of oil palm trees.