A two-step approach for detecting Striga in a complex agroecological system using Sentinel-2 data

Information on weed occurrence within croplands is vital but is often unavailable to support weeding practices and improve cropland productivity assessments. To date, few studies have been conducted to estimate and map weed abundances within agroecological systems from spaceborne images over wide-area landscapes, particularly for the genus Striga. Therefore, this study attempts to increase the detection capacity of Striga at subpixel size using spaceborne high-resolution imagery. In this study, a two-step classification approach was used to detect Striga (Striga hermonthica) weed occurrence within croplands in Rongo, Kenya. Firstly, multidate and multiyear Sentinel-2 (S2) data (2017 to 2018) were utilized to map cropland and non-cropland areas using the random forest algorithm within the Google Earth Engine. The non-cropland class was thereafter masked out from a single date S2 image of the 13th of December 2017. The remaining cropland area was then used in a subpixel multiple endmember spectral mixture analysis (MESMA) to detect Striga occurrence and infestation using endmembers (EMs) obtained from the in-situ hyperspectral data. The gathered in-situ hyperspectral data were resampled to the spectral waveband configurations of S2 and three representative EMs were inferred, namely: (1) Striga, (2) crop and other weeds, and (3) soil. Overall classification accuracies of 88% and 78% for the pixel-based cropland mapping and subpixel Striga detection were achieved, respectively. Furthermore, an F-score (0.84) and a root mean square error (0.0075) showed that the MESMA subpixel algorithm provides plausible results for predicting the relative abundance of Striga within each S2 pixel at a landscape scale. The capability of MESMA together with a cropland classification hierarchical approach was thus proven to be suited for Striga detection in a heterogenous agroecological system. These results can be used to guide in the adaptation, mitigation, and remediation of already infested areas, thereby avoiding further Striga infestation of new croplands. [Display omitted] •African cropping systems hinder the spaceborne detection of Striga weeds•Remotely sensed spectral data often shadows analytical spectral characteristics•A cropland mask was used as fraction constraint in MESMA to enhance Striga detection•This integration successfully reduced spectral confusion and computation time•The method is plausible for Striga detection in complex agroecological systems