Comparative analysis of spectral unmixing methods for subpixel weed identification under controlled and field conditions

Efficient Site-Specific weed management (SSWM) practice requires high-resolution data in both spatial and spectral domains. It allows spatially locating different weed species at early growth to adjust herbicide application based on weed composition and coverage. Nonetheless, such high-resolution data is not always available, and mixed pixels are likely to exist, creating a challenge to generate accurate weed maps. In this regard, Spectral Mixture Analysis (SMA), which allows exploiting subpixel information from coarse spatial resolution spectral data, can mitigate this challenge. This study assesses the potential benefits of four SMA methods for estimating weed coverage from different botanical groups. Each of the methods examined here namely, Fully Constrained Least Squares Unmixing (FCLSU), Sparse Unmixing via variable Splitting and Augmented Lagrangian (SUnSAL), Sparse Unmixing via variable Splitting and Augmented Lagrangian and Total variation (SUnSAL-TV) and the e Vectorized Code Projected Gradient Descent Unmixing (VPGDU) suggests a distinct advantage for spectral unmixing. To compare the four methods, we first established a controlled dataset that included weed species characterized by distinct botanical groups and assessed the performance of four SMA methods in estimating weed coverage and composition at various spatial resolutions. We found that SUnSAL-TV and VPGDU outperformed FCLSU and SUnSAL, with up to 13 % lower Mean Absolute Error (MAE) values. Next, we applied the comparative analysis of these four SMA methods to a multispectral field dataset involving corn and weeds. The same results trend was observed in the field study, with VPGDU as the best-performing method, with an overall MAE value lower than 12 %. These experimental outcomes demonstrate the advantages of the total variation regularization of SUnSAL-TV and the superiority of the SAM-based method, VPGDU, over other approaches, underscoring the advantage of its objective function and the significant effect of varying illumination on the results.