Menu
STARS Project
close
STARS Project
close

Refine your search

STARS Project
Knowledge Portal

Multiple Endmember Spectral Mixture Analysis (MESMA) on multi-temporal VHR images for weed detection in smallholder farms

Need for weed detection

Lack of weed control is a major constraint in crop production, which is a key sector in Sub-Saharan countries like Mali. Remote sensing could play a role in detection and monitoring of weeds. However, crops and weeds occur in mixture. Temporal and spectral information relating to crops and weeds as well as the soil contribution need to be included in a sub-pixel analysis to determine the contribution of weeds, crops and soil in each pixel. Moreover, both the crop and the weeds show a high within class spectral heterogeneity (see Figure 1), because different weeds species may be present within the same field, while the crop and the weeds show at the same time spectral overlap, both being vegetation.

MESMA for weed detection

In this research, the weed fraction is detected based on variations in the spectral response of the plant canopy and depends on the species and the weed density present. Multiple Endmember Spectral Mixture Analysis (MESMA) was used for the detection of weed infestation in maize (Zea mays L.), Cotton (Gossypiumhirsutum) and millet (Pennisetum glaucoma L.). Field boundaries, plots and quadrants were laid out in three small-holder crop fields for the experiment by the STARS project in Sukumba, Mali. Each field had an area of less than two hectare’s. Multi–spectral images consisting of 8 bands obtained by the WorldView-2 satellite platform dating from June to November were used. Spectral libraries were extracted using the Sequential Maximum Angle Convex Cone (SMACC) algorithm, followed by the Spectral Matching Algorithm (SMA) using the Spectral Angle Mapper for the identification of unknown spectra.

Utilization of this information across multiple date’s imagery requires approaches for building correct spectral libraries as well as the use of an accurate spectral mixture classification technique. Due to spectral variability within a class, more than one endmember is required for accurate discrimination between the weeds and crops.

Figure 1: A field where a mixture of crops and weeds causes within field spectral heterogeneity. Source: (STARS&ICRISAT 2014)

MESMA allows multiple endmembers per class. MESMA is an improvement of Simple Linear Spectral Mixture Analysis (SLSMA), which allows only one endmember per class. The key to MESMA is to detect which spectra in a group of spectra are the best representative of a class they represent while covering the range of variability within the class.  Endmember Average Root mean square error (EAR) (Dennison & Roberts, 2003a) and the Minimum Average Spectral Angle (MASA) (Dennison et al., 2004) were used for the endmember selection. A four endmember-model (crop, weed, soil and shadow) was used to un-mix each image independently. The resulting fraction indicated that MASA endmembers modelled the images accurately and outperformed the EAR by having a higher percentage of image pixels modelled and also higher correlations with the ground reference data across all the dates and crop fields. The selected MASA endmembers were used to model the WV-2 images. A shade normalization of the fraction images was performed by dividing each endmember by the total percent of all the non-shade endmembers in each pixel. This suppressed the shade fraction to obtain more information on the relative abundance of non-shade endmembers. Crop and weed maps resulting from MASA for the four endmember model are shown in Figure 2. Red represents the weed class, and green is the crop class. Crop is shown as the dominant class, but weeds are much more widely spread in the November images when the crop is ripening and drying.

  • Figure 2: The MASA endmember classification image per date for the cotton field (left) and the maize field (right). Green represents the crop fraction and red the weed fraction. Figure 2: The MASA endmember classification image per date for the cotton field (left) and the maize field (right). Green represents the crop fraction and red the weed fraction.

Finally, the modelled weed fractions were validated against the weed fractions in the field using root mean square error (RMSE), standard error (SE) and the coefficient of determination (R2). The relationship between the estimated and observed weed densities was satisfactory, the highest coefficient of determination for the cotton field was R2=0.722, with RMSE=0.041 and a SE= 0.003 in November 14th while the lowest was in October18th, with R2=0.591, RMSE=0.043 and SE= 0.001 for the MASA results, which showed a higher agreement than the corresponding EAR results for all dates. These MESMA results showed higher correlation and lower error than the SLMSA results.

Relation between vegetation indices, Fcover and weed fraction

The evaluation of seven vegetation indices for the detection of vegetation (weeds and crops together) showed that the World-View Improved Vegetative Index (WV-VI), which uses NIR2, showed the highest correlation with the fraction of the surface covered by green vegetation (Fcover) measured in the field. High weed fractions coincide with high values of WV-VI (Figure 3).

  • Figure 3: The WV-VI for the cotton field (left) and the corresponding MESMA weed and crop classification image (right) using the MASA modelled fractions. Figure 3: The WV-VI for the cotton field (left) and the corresponding MESMA weed and crop classification image (right) using the MASA modelled fractions.


MESMA modelled the weed fraction more accurately than SLSMA. For both MESMA and SLSMA, MASA endmember selection gave more accurate weed fraction estimates than EAR endmember selection. The use of hyperspectral image data or high-resolution UAV imagery could further improve crop-weed discrimination, because more pure crop- and weed pixels would be available.