By Francisco Arias Rojas
The late blight disease caused by the fungus-like Phytophthora infestans is considered one of the most important and devastating diseases, reducing the potato crop production and affecting the agricultural economy and food security worldwide. Moreover, one of the main characteristics of the late blight disease, is that it spreads extremely quickly during growing season, especially in those production systems where no chemicals are applied in the field, such as those organic farming systems, where production tend to be lower comparing to conventional practices due the constrains on the use of fertilizers and pesticides. The need of real time, location-precise, non-invasive methods to assess late blight disease could provide reliable information for the crop management community to establish suitable solutions for this agricultural practice.
The extensive exploration and use of remote sensing over the past years to retrieve physical variables to estimate vegetation properties, has contributed to the development and improvement of different applications of remotely sensed observations to provide spatial, temporal and non-invasive field-based information relative to plant condition, and consequently the detection, identification, quantification and prediction of crop stress, representing one of the major contributions of this science to agriculture. However, this progress in the remote sensing domain and sensor technologies in combination with UAV platforms, brings together the generation of large amount of data that requires of complex analysis to find underlying information that in most cases statistical approaches are sometimes limited to reach.
Progress in computer science allowed the development and design of new approaches such as machine learning algorithms to analyze in depth the full spectrum offered by the hyperspectral technology to retrieve spectral responses related to plant health status. Even though, most of the studies related to plant disease assessment were applied under controlled conditions (laboratory or greenhouses), some authors have been working intensively to transfer results obtained under controlled environmental conditions to experiments performed at field level. Hence, studies for monitoring fields and plots could provide new insights to better understand the complex host-pathogen relation with the disease spread and its spatial distribution.
Therefore, this study focused to explore the potential use of machine learning techniques for the assessment of late blight disease in an organic potato production system from high-resolution UAV imagery. Moreover, this study explored which discriminant function (linear – non-linear) of a support vector machine classifier provided the best solution and performance to predict late blight disease. The experiment was conducted over eight experimental plots, where two different production systems were evaluated through visual assessments (ground truth), together with the acquisition of high-resolution UAV imagery. The combination of both datasets was used as input to train and evaluate the selected predictive model.
A support vector machine learning algorithm was used to classify three different disease severity classes, which due to the disease evolution over the growing season, labeled classes were represented by an unbalanced overall distribution. As expected, a high correlation between the acquired spectral features was found, therefore during the model selection procedure, a preprocessing feature selection technique such as the PCA was incorporated within the pipeline, to evaluated if the new uncorrelated components contribute for the selection of the best model. Results obtained from the grid search procedure, showed that the incorporation of the first five new components contributed to the selection of the model with the best classification performance, yielding in a balanced accuracy of 82%. In addition, both linear and non-linear SVM discriminant function were tested, to explore which extension of the classifier was more suitable to predict late blight disease. As most of the researches reviewed for this study, among the methods evaluated, a radial basis kernelized SVM algorithm was selected during the model selection procedure.
This study concludes, based on the literature reviewed and the results obtained, that the use of hyperspectral sensors in combination with machine learning techniques such as the support vector machine, has the potential to monitor crop health status. However, it is still necessary to consider factors such as environmental conditions during flight acquisition, image preprocessing and the selection of well-known specific features for the development of automated systems that can provide timely, non-invasive, and reliable information to forecast temporal and spatial disease spread, information that can be use by the crop protection community.