Study of XAI-capabilities for early diagnosis of plant drought

Abstract

The Single Layer Perceptron (SLP) has been studied as an Explainable Artificial Intelligence (XAI) Interactive Unit. On the basis of SLP(N), with an arbitrary number N of neurons on the hidden layer, two models were built: classification and regression. To achieve interactivity, the training on images is replaced by training on its feature vectors. The feature vector includes the results of image processing in three different ways, forming 3 feature groups: STAT {mean, std, min, max}; HIST - values of the quantized histogram; GLCM (gray-level co-occurrence matrix) - textural features. To give XAI properties to the models, they are equipped with tools for analyzing and visualizing the weight and efficiency of the components of the feature vector. It is also possible to optimize the classifier and regressor by the number of neurons, features, and quantization levels (histogram bins and gray levels for GLCM). The study was carried out on the example of the problem of early diagnosis of drought stress in wheat plants, recorded by sensors of two different types: Thermal IR (TIR) and RGB. The problems of stress classification and prediction (regression) of the duration of a plant being under stress are solved. The SLP classifier and the SLP regressor are also used as tools for analyzing the stress features efficiency. Two groups of grayscale NDVI (normalized difference vegetation index) images were used as source data: TIR-based; RGB-based. Replacing source images onto their feature vectors gave to reduce the training time of the models to a fraction of a second. The weights and the influence of drought stress features on the efficiency of classification and regression for both types of source images were shown, and SLP models were optimized. Software tools: pytorch, scikit-image, scikit-learn.