Louis Kouadio

An integrated, probabilistic model for improved seasonal forecasting of agricultural crop yield under environmental uncertainty

We present a novel forecasting method for generating agricultural crop yield forecasts at the seasonal and regional-scale, integrating agroclimate variables and remotely-sensed indices. The method devises a multivariate statistical model to compute bias and uncertainty in forecasted yield at the Census of Agricultural Region (CAR) scale across the Canadian Prairies. The method uses robust variable-selection to select the best predictors within spatial subregions. Markov-Chain Monte Carlo (MCMC) simulation and random forest-tree machine learning techniques are then integrated to generate sequential forecasts through the growing season. Cross-validation of the model was performed by hindcasting/backcasting it and comparing its forecasts against available historical data (1987-2011) for spring wheat (Triticum aestivum L.). The model was also validated for the 2012 growing season by comparing its forecast skill at the CAR, provincial and Canadian Prairie region scales against available statistical survey data. Mean percent departures between wheat yield forecasted were under-estimated by 1-4 % in mid-season and over-estimated by 1 % at the end of the growing season. This integrated methodology offers a consistent, generalizable approach for sequentially forecasting crop yield at the regional-scale. It provides a statistically robust, yet flexible way to concurrently adjust to data-rich and data-sparse situations, adaptively select different predictors of yield to changing levels of environmental uncertainty, and to update forecasts sequentially so as to incorporate new data as it becomes available. This integrated method also provides additional statistical support for assessing the accuracy and reliability of model-based crop yield forecasts in time and space.

Estimating regional wheat yield from the shape of decreasing curves of green area index temporal profiles retrieved from MODIS data

Earth observation data, owing to their synoptic, timely and repetitive coverage, have been recognized as a valuable tool for crop monitoring at different levels. At the field level, the close correlation between green leaf area (GLA) during maturation and grain yield in wheat revealed that the onset and rate of senescence appeared to be important factors for determining wheat grain yield. Our study sought to explore a simple approach for wheat yield forecasting at the regional level, based on metrics derived from the senescence phase of the green area index (GAI) retrieved from remote sensing data. This study took advantage of recent methodological improvements in which imagery with high revisit frequency but coarse spatial resolution can be exploited to derive crop-specific GAI time series by selecting pixels whose ground-projected instantaneous field of view is dominated by the target crop: winter wheat. A logistic function was used to characterize the GAI senescence phase and derive the metrics of this phase. Four regression-based models involving these metrics (i.e., the maximum GAI value, the senescence rate and the thermal time taken to reach 50% of the green surface in the senescent phase) were related to official wheat yield data. The performances of such models at this regional scale showed that final yield could be estimated with an RMSE of 0.57 ton ha−1, representing about 7% as relative RMSE. Such an approach may be considered as a first yield estimate that could be performed in order to provide better integrated yield assessments in operational systems.

Assessing the Performance of MODIS NDVI and EVI for Seasonal Crop Yield Forecasting at the Ecodistrict Scale

Crop yield forecasting plays a vital role in coping with the challenges of the impacts of climate change on agriculture. Improvements in the timeliness and accuracy of yield forecasting by incorporating near real-time remote sensing data and the use of sophisticated statistical methods can improve our capacity to respond effectively to these challenges. The objectives of this study were (i) to investigate the use of derived vegetation indices for the yield forecasting of spring wheat (Triticum aestivum L.) from the Moderate resolution Imaging Spectroradiometer (MODIS) at the ecodistrict scale across Western Canada with the Integrated Canadian Crop Yield Forecaster (ICCYF); and (ii) to compare the ICCYF-model based forecasts and their accuracy across two spatial scales-the ecodistrict and Census Agricultural Region (CAR), namely in CAR with previously reported ICCYF weak performance. Ecodistricts are areas with distinct climate, soil, landscape and ecological aspects, whereas CARs are census-based/statistically-delineated areas. Agroclimate variables combined respectively with MODIS-NDVI and MODIS-EVI indices were used as inputs for the in-season yield forecasting of spring wheat during the 2000–2010 period. Regression models were built based on a procedure of a leave-one-year-out. The results showed that both agroclimate + MODIS-NDVI and agroclimate + MODIS-EVI performed equally well predicting spring wheat yield at the ECD scale. The mean absolute error percentages (MAPE) of the models selected from both the two data sets ranged from 2% to 33% over the study period. The model efficiency index (MEI) varied between −1.1 and 0.99 and −1.8 and 0.99, respectively for the agroclimate + MODIS-NDVI and agroclimate + MODIS-EVI data sets. Moreover, significant improvement in forecasting skill (with decreasing MAPE of 40% and 5 times increasing MEI, on average) was obtained at the finer, ecodistrict spatial scale, compared to the coarser CAR scale. Forecast models need to consider the distribution of extreme values of predictor variables to improve the selection of remote sensing indices. Our findings indicate that statistical-based forecasting error could be significantly reduced by making use of MODIS-EVI and NDVI indices at different times in the crop growing season and within different sub-regions.

Brown rust disease control in winter wheat: II. Exploring the optimization of fungicide sprays through a decision support system

A decision support system (DSS) involving an approach for predicting wheat leaf rust (WLR) infection and progress based on night weather variables (i.e., air temperature, relative humidity, and rainfall) and a mechanistic model for leaf emergence and development simulation (i.e., PROCULTURE) was tested in order to schedule fungicide time spray for controlling leaf rust progress in wheat fields. Experiments including a single fungicide treatment based upon the DSS along with double and triple treatment were carried out over the 2007–2009 cropping seasons in four representative Luxembourgish wheat field locations. The study showed that the WLR occurrences and severities differed according to the site, cultivar, and year. We also found out that the single fungicide treatment based on the DSS allowed a good protection of the three upper leaves of susceptible cultivars in fields with predominant WLR occurrences. The harvested grain yield was not significantly different from that of the double and triple fungicide-treated plots (P < 0.05). Such results could serve as basis or be coupled to cost-effective and environmentally friendly crop management systems in operational context.

Disease Severity Estimates – Effects of Rater Accuracy and Assessment Methods for Comparing Treatments

Assessment of disease severity is required for several purposes in plant pathology; most often, the estimates are made visually. It is established that visual estimates can be inaccurate and unreliable. The ramifications of biased or imprecise estimates by raters have not been fully explored using empirical data, partly because of the logistical difficulties involved in different raters assessing the same leaves for which actual disease has been measured in a replicated experiment with multiple treatments. In this study, nearest percent estimates (NPEs) of Septoria leaf blotch (SLB) on leaves of winter wheat from nontreated and fungicide-treated plots were assessed in both 2006 and 2007 by four raters and compared with assumed actual values measured using image analysis. Lins concordance correlation (LCC, ρc) was used to assess agreement between the two approaches. NPEs were converted to Horsfall-Barratt (HB) midpoints and were compared with actual values. The estimates of SLB severity from fungicide-treated and nontreated plots were analyzed using generalized linear mixed modeling to ascertain effects of rater using both the NPE and HB values. Rater 1 showed good accuracy (ρc = 0.986 to 0.999), while raters 3 and 4 were less accurate (ρc = 0.205 to 0.936). Conversion to the HB scale had little effect on bias but reduced numerically both precision and accuracy for most raters on most assessment dates (precision, r = -0.001 to -0.132; and accuracy, ρc = -0.003 to -0.468). Interrater reliability was also reduced slightly by conversion of estimates to HB midpoint values. Estimates of mean SLB severity were significantly different between image analysis and raters 2, 3, and 4, and there were frequently significant differences among raters (F = 151 to 1,260, P = 0.001 to P < 0.0001). Only on 26 June 2007 did conversion to the HB scale change the means separation ranking of rater estimates. Nonetheless, image analysis and all raters were able to differentiate control and treated-plot treatments (F = 116 to 1,952, P = 0.002 to P < 0.0001, depending on date and rater). Conversion of NPEs to the HB scale tended to reduce F values slightly (2006: NPEs, F = 116 to 276, P = 0.002 to 0.0005; and, for the HB-converted values, F = 101 to 270, P = 0.002 to 0.0005; 2007: NPEs, F = 164 to 1,952, P = 0.001 to P < 0.0001; and, for HB-converted values, F = 126 to 1,633, P = 0.002 to P < 0.0001). The results reaffirm the need for accurate and reliable disease assessment to minimize over- or underestimates compared with actual disease, and the data we present support the view that, where multiple raters are deployed, they should be assigned in a manner to reduce any potential effect of rater differences on the analysis.

A threshold-based weather model for predicting stripe rust infection in winter wheat

Wheat stripe rust (caused by Puccinia striiformis f. sp. tritici) is a major threat in most wheat growing regions worldwide, which potentially causes substantial yield losses when environmental conditions are favorable. Data from 1999 to 2015 for three representative wheat-growing sites in Luxembourg were used to develop a threshold-based weather model for predicting wheat stripe rust. First, the range of favorable weather conditions using a Monte Carlo simulation method based on the Dennis model were characterized. Then, the optimum combined favorable weather variables (air temperature, relative humidity, and rainfall) during the most critical infection period (May-June) was identified and was used to develop the model. Uninterrupted hours with such favorable weather conditions over each dekad (i.e., 10-day period) during May-June were also considered when building the model. Results showed that a combination of relative humidity >92% and 4°C < temperature < 16°C for a minimum of 4 continuous hours, associated with rainfall ≤0.1 mm (with the dekad having these conditions for 5 to 20% of the time), were optimum to the development of a wheat stripe rust epidemic. The model accurately predicted infection events: probabilities of detection were ≥0.90 and false alarm ratios were ≤0.38 on average, and critical success indexes ranged from 0.63 to 1. The method is potentially applicable to studies of other economically important fungal diseases of other crops or in different geographical locations. If weather forecasts are available, the threshold-based weather model can be integrated into an operational warning system to guide fungicide applications.