Understanding plant health status from electrical impedance using Neural Networks

This thesis is about an exploration in the world of machine learning to find a way to predict the health state of a plant, in particular tobacco plants. The prediction is made possible by using an intrinsic measurement of the plant, the electrical impedance, and the surrounding environment, such as temperature, air humidity, soil moisture, ambient light, and the time when the measurements are taken. Predicting the plant status by knowing stem impedance, module and phase, and the environmental measurements can be seen as a classification problem. Therefore, the most suitable machine learning algorithms are the supervised learning ones. The most common supervised learning algorithms are the k-Nearest neighbor algorithm, Decision Tree, logistic regression, Support Vector Machine, and Neural Networks. Finally, Artificial Neural Networks are chosen for this work for many reasons, such as their ability to learn and model non-linear and complex relationships; they do not impose any restrictions on the input variables. Preparing the dataset for the training is one of the most challenging parts, not only because the sampled data, the inputs, must make sense, but also the labels, i.e., the outputs, must be correct. In order to label the plants correctly, the pictures of the plants in this period are looked at one by one, and the health status is determined based on the leaves conditions of the tobacco plants. Once the dataset is ready, the neural network has been implemented in Python by exploiting the open-source library TensorFlow. The performance reached is 82.5% of training accuracy and 81.3% of testing accuracy. Furthermore, this model is used to predict the outputs of future data, reaching 90% of total accuracy on the four plants. Then three plants are used as training set, and the fourth is used for the testing. However, the results are not very promising. The bad performance is mainly due to each plant's different impedance module range for both the first and subsequent networks. Two different methods are tried to solve this problem. The first one considers the impedance difference between two adjacent samples as an additional feature for the training. The second one is to consider the difference after a polynomial fitting of the data, filtering out the daily cycle variation and considering only the general trend of the impedance. Nevertheless, unfortunately, in both cases, the accuracy remains similar to the first implementation. Finally, a neural network considering the past values of module and phase as features is implemented. Considering 48 times the module and phase, which means 48 hours of samples, the neural network reaches a training accuracy of 98.3% and testing accuracy of 96.6%. This type of network also has excellent performances for future prediction. This approach is also applied to the training with three plants and testing with the fourth one, but the performances are awful, and the problem of different modules' value for different plants is still unsolved. This thesis work demonstrates the importance of impedance and soil moisture for detecting a tobacco health status; this result is confirmed by a simple SVM implementation with a radial basis function kernel. The neural network has reached significant results, but it is still not ready to be used on an unknown tobacco plant, which means not used during the training phase.