Detection, identification, and mapping of maize streak virus and grey leaf spot diseases of maize using different remote sensing techniques

Abstract

Of late climate change and consequently, the spread of crop diseases has been identified as one of the major threat to crop production and food security in subSaharan Africa. This research, therefore, aims to evaluate the role of in situ hyperspectral and new generation multispectral data in detecting maize crop viral and fungal diseases, that is maize streak virus and grey leaf spot respectively. To accomplish this objective; a comparison of two variable selection techniques (Random Forest’s Forward Variable, (FVS) and Guided Regularized Random Forest: (GRRF) was done in selecting the optimal variables that can be used in detecting maize streak virus disease using in-situ resampled hyperspectral data. The findings indicated that the GRRF model produced high classification accuracy (91.67%) whereas the FVS had a slightly lower accuracy (87.60%) based on Hymap when compared to the AISA. The results have shown that the GRRF algorithm has the potential to select compact feature sub sets, and the accuracy performance is better than that of RF’s variable selection method. Secondly, the utility of remote sensing techniques in detecting the geminivirus infected maize was evaluated in this study based on experiments in Ofcolaco, Tzaneen in South Africa. Specifically, the potential of hyperspectral data in detecting different levels of maize infected by maize streak virus (MSV) was tested based on Guided Regularized Random Forest (GRRF). The findings illustrate the strength of hyperspectral data in detecting different levels of MSV infections. Specifically, the GRRF model was able to identify the optimal bands for detecting different levels of maize streak disease in maize. These bands were allocated at 552 nm, 603 nm, 683 nm, 881 nm, and 2338 nm. This study underscores the potential of using remotely sensed data in the accurate detection of maize crop diseases such as MSV and its severity which is critical in crop monitoring to foster food security, especially in the resource-limited subSaharan Africa. The study then investigated the possibility to upscale the previous findings to space borne sensor. RapidEye data and derived vegetation indices were tested in detecting and mapping the maize streak virus. The results revealed that the use of RapidEye spectral bands in detection and mapping of maize streak virus disease yielded good classification results with an overall accuracy of 82.75%. The inclusion of RapidEye derived vegetation indices improved the classification accuracies by 3.4%. Due to the cost involved in acquiring commercial images, like xviii

RapidEye, a freely available Landsat-8 data can offer a new data source that is useful for maize diseases estimation, in environments which have limited resources. This study investigated the use of Landsat 8 and vegetation indices in estimating and predicting maize infected with maize streak virus. Landsat 8 data produced an overall accuracy of 50.32%. The inclusion of vegetation indices computed from Landsat 8 sensor improved the classification accuracies by 1.29%. Overally, the findings of this study provide the necessary insight and motivation to the remote sensing community, particularly in resource-constrained regions, to shift towards embracing various indices obtained from the readily-available and affordable multispectral Landsat-8 OLI sensor. The results of the study show that the mediumresolution multispectral Landsat 8-OLI data set can be used to detect and map maize streak virus disease. This study demonstrates the invaluable potential and strength of applying the readily-available medium-resolution, Landsat-8 OLI data set, with a large swath width (185 km) in precisely detecting and mapping maize streak virus disease. The study then examined the influence of climatic, environmental and remotely sensed variables on the spread of MSV disease on the Ofcolaco maize farms in Tzaneen, South Africa. Environmental and climatic variables were integrated together with Landsat 8 derived vegetation indices to predict the probability of MSV occurrence within the Ofcolaco maize farms in Limpopo, South Africa. Correlation analysis was used to relate vegetation indices, environmental and climatic variables to incidences of maize streak virus disease. The variables used to predict the distribution of MSV were elevation, rainfall, slope, temperature, and vegetation indices. It was found that MSV disease infestation is more likely to occur on low-lying altitudes and areas with high Normalised Difference Vegetation Index (NDVI) located at an altitude ranging of 350 and 450 m.a.s.l. The suitable areas are characterized by temperatures ranging from 24°C to 25°C. The results indicate the potential of integrating Landsat 8 derived vegetation indices, environmental and climatic variables to improve the prediction of areas that are likely to be affected by MSV disease outbreaks in maize fields in semi-arid environments. After realizing the potential of remote sensing in detecting and predicting the occurrence of maize streak virus disease, the study further examined its potential in mapping the most complex disease; Grey Leaf Spot (GLS) in maize fields using WorldView-2, Quickbird, RapidEye, and Sentinel-2 resampled from hyperspectral data. To accomplish this objective, field spectra were acquired from healthy, moderate and xix

severely infected maize leaves during the 2013 and 2014 growing seasons. The spectra were then resampled to four sensor spectral resolutions – namely WorldView-2, Quickbird, RapidEye, and Sentinel-2. In each case, the Random Forest algorithm was used to classify the 2013 resampled spectra to represent the three identified disease severity categories. Classification accuracy was evaluated using an independent test dataset obtained during the 2014 growing season. Results showed that Sentinel-2 achieved the highest overall accuracy (84%) and kappa value (0.76), while the WorldView-2, produced slightly lower accuracies. The 608 nm and 705nm were selected as the most valuable bands in detecting the GLS for Worldview 2, and Sentinel-2. Overall, the results imply that opportunities exist for developing operational remote sensing systems for detection of maize disease. Adoption of such remote sensing techniques is particularly valuable for minimizing crop damage, improving yield and ensuring food security.