Prediction of Ecophysiological Variables from Remote Sensing Data Using Machine Learning Methods

Description

The project makes use of a dataset consisting of visible and near-infrared (NIR) reflectance values of maize and soybean leaves, as well as the corresponding leaf-level measurements of several ecophysiological variables. The reflectance values were recorded with a hyperspectral camera and the leaf-level measurements were made with handheld devices. The dataset also contains a set of top view thermal images of the plants. This thesis project is organised in two parts. Firstly, plant temperature values were automatically extracted from the thermal images. This was achieved by an initial exploration in the field of image analysis, which resulted in the proposal of three temperature extraction methods. These methods were then evaluated and compared on a supervised test set of 15 thermal images. Secondly, taking the average temperature value, the standard deviation of the temperature, and the hyperspectral reflectance data, the modeling capabilities of multi-block partial least squares regression models were evaluated for stomatal conductance, photosynthetic rate, and transpiration. When predicting stomatal conductance, further analysis was performed to investigate the reasons behind the found performance. In addition, a final step was to quantify the importance of the thermal information as an extension to the hyperspectral data.

Period	1 Jul 2019 → 31 Dec 2019
Examinee
Examination held at	Department of Applied Mathematics and Computer Science
Degree of Recognition	Local