Digital camera operation and fuzzy logic classification of uniform plant, soil, and residue color images

Low-cost, consumer grade cameras have been used in recent machine vision studies. Users are faced with a choice of manual or automatic operations for obtaining quality images for classifying plant, soil, and residue for field maps and precision agriculture. A digital camera operations study was conducted for classifying uniform images of grass, bare soil, corn stalks residue, wheat straw residue, and a barium sulfate reference panel, based on color. Both natural and artificial background lighting was studied. Classifications were conducted with fuzzy inference systems, built with subtractive clustering, an Adaptive-Network Fuzzy Inference System (ANFIS) and ten-fold cross-validation. Each image was labeled with an integer for classification and reference. Classification systems contained from 5 to 36 rules. Rules for automatic digital camera operation were well trained (r² > 0.94), with low root-mean-squae errors (RMSE < 0.2). Manual digital camera operation gave lower training correlations (r² < 0.7) and higher root-mean-squae errors (RMSE > 3). Since predicted classification values were not integral, Zadeh's principle of maximum membership was used for final classification using trained output membership functions. Correct classification rates (CCR) for manual camera operation were low (<64%), and only slightly improved by eliminating over- and underexposed and monochromatic-lit images. The automatic digital camera provided better classification rates (>>81%). Independent background lighting luminosity and color temperature measurements did not significantly improve classification. The results suggest that low-cost, digital cameras used in automatic mode would be best for remote sensing and site-specific crop management use.