Identification of Weed Species and Glyphosate-Resistant Weeds Using High Resolution UAS Images

Adoption of a Site-Specific Weed Management System (SSWMS) can contribute to sustainable agriculture. Weed mapping is a crucial step in SSWMS, leads to saving herbicides and protecting environment by preventing repeated chemical applications. In this study, the feasibility of visible and near infrared spectroscopy to classify three problematic weed species and to identify glyphosate-resistant weeds was evaluated. The canopy temperature was also employed to identify the glyphosate-resistant weeds. Furthermore, the ability of UAS imagery to develop accurate weed map in early growing season was evaluated. A greenhouse experiment was conducted to classify waterhemp (Amaranthus rudis), kochia (Kochia scoparia), and lambsquartes (Chenopodium album) based on spectral signature. The Soft Independent Modeling of Class Analogy (SIMCA) method on NIR (920-2500 nm) and Vis/NIR (400-2500 nm) regions classified three different weed species with accuracy greater than 90 %. The discrimination power of different wavelengths indicated that 640, 676, and 730 nm from red and red-edge region, and 1078, 1435, 1490, and 1615 nm from the NIR region was the best wavelengths for weed species discrimination. While, wave 460, 490, 520 and 670 nm from Vis range, and 760, 790 nm from NIR region were the significant discriminative features for identifying glyphosate-resistant weeds. Random Forest was able to detect glyphosate-resistant weeds based on spectral weed indices with more than 95% accuracy. Analysis of thermal images indicated that the canopy temperature of glyphosate-resistant weeds was less than susceptible ones early after herbicide application. The test set validation results showed the support vector machine method could classify resistant weed species with accuracy greater than 95 %. Based on the stepwise method the best times for discrimination of kochia, and waterhemp resistant were 46 and 95 hours after glyphosate application, respectively. In addition, a field study was proposed on soybean field to identify weed species and glyphosate-resistant weeds using multispectral and thermal imagery. Results revealed that the object-based supervised classification method could classify weed species with greater than 90% accuracy in early growing season. Furthermore, the glyphosate-resistant kochia, waterhemp and ragweed were identified based on canopy temperature with 88%, 93% and 92% accuracy, respectively