Coconut trees are an important part of the tropical islands' agricultural systems which are increasingly prone to natural disasters, making frequent monitoring of this natural resource critical to assessments of potential or actual impacts of catastrophic events. NASA Harvest researchers Sergii Skakun [UMD] and Inbal Becker-Reshef [NASA Harvest Program Director] in collaboration with partners at the NASA Goddard Space Flight Center (E.Vermote) and World Bank (K.Saito) recently published a paper on "Remote Sensing of Coconut Trees in Tonga Using Very High Spatial Resolution WorldView-3 Data" in the journal Remote Sensing. This study used very high spatial resolution satellite imagery (at 30 cm resolution) acquired by WordView-3 to count coconut trees in the Tonga archipelago, with the resulting algorithm proving useful for estimating the impact of natural disasters, such as 2018 cyclone Gita, on the region's coconut plantations.
This paper is the first to be published in the Special Issue "Remote Sensing of Land Use/Cover Changes Using Very High Resolution Satellite Data" that aims at NASA-funded researchers who have been using very high resolution data in land use and crop cover research and applications. Earth observation data from satellites in space offer capabilities to regularly and spatially monitor Earth’s resources in great detail. With advances in machine learning methods, such as deep learning, much effort has been put into applying very high resolution agricultural monitoring methods for tree detection via satellite imagery. However, proper training of deep learning models necessitates a high volume of labeled training data, which is time-consuming, complicated, and not always possible when damages need to be quickly assessed after a natural disaster occurs. In this paper, the team of researchers present an automated approach for counting coconut trees with an algorithm tailored towards coconut plantations that are chiefly operating for agricultural purposes, applied over the Tonga region.
This paper presents a simple and efficient image processing method for estimating the number of coconut trees in the Tonga region using very high spatial resolution data (30 cm) in the blue, green, red and near infrared spectral bands acquired by the WorldView-3 sensor. The method is based on the detection of tree shadows and the further analysis to reject false detection using geometrical properties of the derived segments. The algorithm is evaluated by comparing coconut tree counts derived by an expert through photo-interpretation over 57 randomly distributed (4% sampling rate) segments of 200 m × 200 m over the Vaini region of the Tongatapu island. The number of detected trees agreed within 5% versus validation data. The proposed method was also evaluated over the whole Tonga archipelago by comparing satellite-derived estimates to the 2015 agricultural census data—the total tree counts for both Tonga and Tongatapu agreed within 3%.
For more details, check out the full publication which is openly available online.