Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Accueil du site → Master → Israel → Estimation of plant height and growth rate in cotton fields from multi-spectral satellite imagery using machine learning methods

Ben-Gurion University of the Negev (2021)

Estimation of plant height and growth rate in cotton fields from multi-spectral satellite imagery using machine learning methods

Makler, Lev

Titre : Estimation of plant height and growth rate in cotton fields from multi-spectral satellite imagery using machine learning methods

הערכת גובה צמח וקצב צימוח בשדות כותנה מתוך דימותי לוויין מולטי-ספקטראליים בשיטות של לימוד מכונה.

Auteur : Makler, Lev,

Etablissement de soutenance : Ben-Gurion University of the Negev

Grade : Master of Science (MS) 2021

Résumé partiel
Irrigation is a significant factor that determines the quantity and quality of crops. Farmers determine the amount of water and dates according to pre-determined recommendations and/or according to several physiological measurements. The irrigation is done uniformly to the whole field, regardless of its spatial variability. In Israel, the common measure for cotton to correct irrigation quantity during the vegetative growth stage is plant height. For that, farmers measure the height of a few plants in a representative area in the field 1-2 times a week. This method is reliable but does not necessarily characterize the field’s heterogeneity. Additionally, because of its laborious nature, the frequency of the measurements is often not optimal. With technological advancement, the availability of free satellite imagery with high spectral, spatial, and temporal resolutions is greater than ever. In many studies, the examination of spectral indices in the evaluation of plant parameters has been done in small-scale experiments. The limitations of such an approach lie in the small scale from which the data is collected, which casts doubt on the generalizability of the proposed models, that is, the extent to which these models are valid in other areas. The current study describes an observational approach and used data from different fields at non-experiment conditions (commercial fields) located in different parts of Israel, collected by farmers using a dedicated app and by-passing manual records. It assumes that using this kind of data, relationships between spectral indices and plant parameters can be more robust with no location dependency. Data was collected for the 2019 and 2020 cotton growing seasons from 22 farms and 151 cotton fields, resulting in a database of 2899 height records. The plot geographical layer, the height records, and spectral vegetation indices (SVIs) from Sentinel-2 and Venµs satellite images were joint spatially and temporally in a python environment to examine various height estimation models. Various SVI time-series were extracted using Google Earth Engine (GEE) environment. Due to the fact that the height data is observational and that remote sensing data is sensitive to noise, a large percentage of the data suffered from various noises like miss-location problems and cloudiness. Therefore, a noise reduction phase must be incorporated in the algorithm developments process utilizing the observational approach. The study examined 16 models covering both multiple linear regressions and machine learning methods (random forest – RF and artificial neural network - ANN) and also a hybrid model that combines stepwise regression as a preliminary step before applying random forest.

Présentation et version intégrale (PRIMO)

Page publiée le 11 janvier 2022