The herbicide 2,4-dichlorophenoxyacetic acid (2,4-D) is one of the most successful selective herbicides used in agriculture to control broadleaf weeds. Unfortunately, cotton crops are highly susceptible to 2,4-D, and they are often damaged by the off- target movement of the active ingredient when sprayed as a herbicide on surrounding farms. This action, referred to as herbicide drift, affects the cotton industry every season, causing losses of millions of dollars. Although the economic repercussions on the industry are high, the traditional (visual) assessment of damage is often imprecise and inaccurate. Crop sensing tools can offer alternative and reliable methods to overcome the typical limitations of visual assessments by providing accurate estimations of crop performance. The aim of this research project was to assess the capabilities of crop sensing techniques of providing spatial and quantitative information of cotton yield after being affected by 2,4-D herbicide drift. This information is valuable to agronomic planners for evaluating their crop management strategies in order to maximise cotton production while safeguarding the environment in the affected area.
The research area was located in a cotton-growing region in Jondaryan, Queensland, Australia. Two study cases and three remote/proximal sensing approaches were tested. The first study case consisted of controlled doses to simulate accidental exposure to 2,4-D, where three doses (D) and three timing of exposures (S) were examined at four different dates after the exposure (DAE): 2, 7, 14 and 28 DAE. In this case, a hyperspectral sensor and a terrestrial laser scanner (TLS) were evaluated to assess their ability to predict yield loss, dose and canopy structure variability. The second case examined the potential capabilities of satellite imagery for yield loss assessment in an uncontrolled exposure of cotton crops to 2,4-D. For this case, several multispectral (Landsat 8 Operational Land Imager - OLI) images were analysed and a comprehensive approach was developed to overcome the potential limitation of moderate resolution imagery at the field level.
The controlled case revealed that hyperspectral data can be used to predict yield loss with high accuracy (R2 = 0.88) regardless of the timing of exposure and dose, and that 7 DAE and 28 DAE (RMSECV: 2.6 bales/ha; R2 = 0.88 and RMSECV: 3.2 bales/ha;
R2 = 0.84, respectively) were the best times for data collection purposes. The main difference in the model performance between the best (7 DAE) and the worst (14 DAE) prediction model was the inclusion of the NIR range, as the 14 DAE was the only model with no significant wavelengths in this range.
Through this case, it was possible to better understand how the internal changes of the contaminated leaves, that is photosynthesis, stomatal conductance and hormone contents, influenced their spectral response and the lint quality of the cotton. Most of the variables analysed in this study manifested a significant relationship with hyperspectral data (p-value < 0.05). The harvested yield was severely affected by the herbicide, with losses recorded as high as 98%, while the fibre quality remained relatively unaffected. The prediction capabilities for the simulated dose were also tested by implementing Canonical Powered PLS (CPPLS) and Sparse PLS Discriminant Analysis (sPLS-DA). High accuracies (> 70%) were obtained regardless of the method, D or S. However, the timing of exposure (S) resulted in being a determinant to improve the classification accuracy to more than 90%.
The analysis of laser scanner-derived data provided accurate information about the canopy height and canopy volume that could be strongly correlated (r > 0.88) with yield at different times of assessment (2 DAE, 7 DAE and 14 DAE). High R2 (> 0.90) between measured and estimated canopy height validates the height values estimated from the TLS-derived data. Furthermore, the weak relationship (R2 =0.39, p-value > 0.05) between point density and estimated canopy volume provided an insight that the approach implemented to estimate cotton canopy height and volume overcame the reported limitations of terrestrial laser scanners in the field.
The uncontrolled case (i.e. Landsat 8 imagery) tested six different dates for optimal data collection purposes. The results demonstrated that traditional vegetation indices
(VI) and individual multispectral bands were incapable of predicting yield in neither affected nor unaffected cotton areas (R2 < 0.27). However, PLS-R models optimised the information provided by the multispectral bands. As a result, the R2 increased, in some cases, by more than 60%. From the PLS- model results, it was determined that one week after the exposure was the best time for the prediction of yield in affected areas (RMSEP = 1.19 bales/ha and R2 = 0.60). Satellite imagery could be then implemented to support targeted monitoring programs in 2,4-D-injured areas.
The technologies implemented in this study were proven to be reliable for damage assessment after an accidental spray drift by accurately predicting yield and dose and also by estimating canopy structure variables strongly correlated with yield in 2,4-D- affected areas. These comprehensive analytical approaches also provided information on temporal windows for optimal data collection after an incident, and also on less- recommended dates for the same purpose. These methods indicated an optimal window between seven and 14 days, or more than 28 days after the exposure, for the prediction of damage. However, as soon as two days after the cotton plant was exposed, hyperspectral measurements and TLS-derived data recorded significant differences in comparison with unaffected control plants.
