Herbicide injury on weeds is quantified through responses such as mortality, biomass, and visual scores, which are time consuming and/or subjective, making it hard to evaluate injury in experiments and making the site-specific management impracticable in major crops. The objective of our work was to build prediction models of glyphosate injury on weed species of economic importance in the Brazilian Cerrado (savana) through multispectral digital images. Two weed species with different levels of susceptibility to glyphosate, Eleusine indica (L.) Pers. and Brachiaria decumbens L., were cultivated in pots and subjected to increasing doses of glyphosate, simulating a gradient of injury. Visual scores were given by three trained evaluators during three weeks after the application. Images of the pots were acquired with a digital multispectral camera in the following bands of the electromagnetic spectrum: blue (450 nm), green (560 nm), red (660 nm), and near-infrared (850 nm), with a digital camera at 0.9-m height from the ground. Images were first segmented to separate plant from soil using the hue–saturation–value color space. Vegetation indices were calculated and used as explanatory variables for phytotoxicity in regression models. A field experiment was conducted in order to validate predictions. The injury pattern of glyphosate on E. inidica and B. decumbens was best modeled by an exponential function of the median of hue for plant pixels, which could be reproduced at field scale with prediction errors below 8%. Differences in sensitivity to glyphosate between the two species were detected with the kernel density of hue for plant pixels, with absolute correlation with phytotoxicity above 0.8.