For accurately isolating the overlapping spots in crop disease images, an improved watershed algorithm named marker-reconstruction watershed algorithm was presented. Firstly, the binary spots image was transformed to distance image by multi-scale distance transform operator and makers (seed) were detected by using ultimate erosion. Then, geodesic SKIZ (i.e., watershed division lines) has been obtained by geodesic reconstruction operator constrained by marking seed image. Finally, the watershed division lines were subtracted from the original binary image. Applying the proposed method in cotton disease images, the experimental results indicated that the method could validly and effectively separate overlapping spots in the image, moreover, preferably protect the spots edge. This method is insensitive to noise and tiny outlier of the image.