Surveying coconut trees using high-resolution satellite imagery in remote atolls of the Pacific Ocean

Coconut (Cocos nucifera L.) is one of the world's most economically important tree species, and coconut palm plantations dominate many islands and tropical coastlines. However, the expansion of plantations to supply international markets threatens biodiversity. Therefore, monitoring the plantations is important not only for the food industry but also for evaluating and mitigating environmental impacts of the industry. However, the detection of coconut trees from space is challenging because the palms' crowns hold only limited pixels of high-resolution optical imagery. Here, we present an accurate and real-time COCOnut tree DETection method (COCODET) which uses satellite imagery to detect individual palms, comprising three components. First, an Adaptive Feature Enhancement (AFE) module is designed to improve both the capacity of representation at the highest level of the feature map and feature representation ability and help distinguish between coconut trees and other vegetation. Secondly, we modify a region proposal network to produce a Tree-shape Region Proposal Network (T-RPN) for producing coconut tree candidates. Finally, we create a Cross Scale Fusion (CSF) module for integrating multi-scale information to improve small tree detection; this fuses features of coconut crowns from different levels, connecting shallow and deep-level semantic features. We applied COCODET to detect coconut trees in four remote atolls from the Acteon Group in French Polynesia. The natural habitats on the islands were previously cleared for coconut plantations, many of which have since been abandoned. COCODET achieved an average F1 score of 86.5% using its real-time inference process, considerably outperforming other cutting-edge object detection algorithms (4.3 ∼ 12.0% more accurate). We detected 688 ha of coconuts and 182 ha of natural habitat on the islands, and within the coconut groves we detected 120,237 individuals. Our analyses indicate that deep learning approaches can be successfully applied to coconut palm detection, aiding efforts to understand human impacts on natural ecosystems and biodiversity.