Invasive plants detection and distribution patterns analysis through Self-Attention enhanced semantic segmentation in UAV imagery and Moran's Index

The development of sustainable agriculture necessitates the rapid identification and efficient removal of invasive plants. Traditionally, the investigation of invasive plants relies on manual surveys, which are prone to subjective errors and require significant human labor. This study introduces an innovative approach for accurately identifying and efficiently eliminating invasive plants. High-resolution aerial images were captured using drones, and a novel semantic segmentation network, based on DeepLabV3+ and the Self-Attention mechanism, was developed to reliably identify two globally distributed invasive species: Erigeron annuus (L.) Pers. and Erigeron canadensis L. at the pixel level. Experimental results from orchard imagery revealed that the proposed approach demonstrated remarkable performance, achieving a mean Precision (mPrecision) of 0.972, a mean Intersection over Union (mIoU) of 0.947, and a mean Pixel Accuracy (mPA) of 0.973 for its overall effectiveness. A high-resolution species distribution map at pixel-level was generated using the results of this model. The study further explored ecological analysis methods of this distribution map, and successfully calculated the coverage area, coverage rate, global Moran's index, and local Moran's index of the two invasive plants. The findings revealed that the intraspecific distribution patterns of both species are characterized by clustering, with Global Moran's Index values of 0.03 for Erigeron annuus (L.) Pers. and 0.23 for Erigeron canadensis L. The clustering map facilitates the rapid identification of invasive plant cluster centers, enabling more targeted weed control measures. The efficient pixel-level invasive plant identification model and species distribution pattern analysis proposed in this study holds significant implications for agricultural production and ecological surveys. They support precise and rapid invasive plant control and to reduce pesticide use. The proposed method can also be implemented on other platforms to provide fast, flexible, and accurate invasive plant mapping and precision agriculture applications.