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

Titre : Remote sensing of croplands, crop productivity, and water use efficiency with a focus on smallholder systems in southern Africa

Auteur : Wellington, Michael

Université de soutenance : Australian National University

Grade : Doctor of Philosophy (PhD) 2023

Description partielle
The increasing availability of large geospatial datasets has allowed development of remote monitoring methods for broadacre crops. Ongoing deployment of new satellite missions means researchers have an increasing volume of data from which information on crop production can be derived. However, more accurate mapping of croplands, estimation of crop productivity, and monitoring of production trends are required. Accommodating the dynamic, seasonal nature of farming systems in mapping and productivity models, and production trend monitoring methods, may enhance their accuracy. Further, these methods must be adapted and applied to diverse farming systems such as smallholder irrigation schemes in southern Africa, and broadacre grain regions in Australia. This thesis deals with several aspects of geospatial analysis of croplands including mapping by image classification, estimation of crop Gross Primary Productivity, and quantifying spatio-temporal trends in crop production and water use efficiency. Mapping methods were developed using known irrigation sites in Zimbabwe, while ground data constraints meant that well-studied global study sites were used to develop the Gross Primary Productivity and trend analysis methods. Finally, these methods were combined to analyse trends for study sites in Mozambique, Tanzania, and Zimbabwe. In each of these analyses, models which captured temporal dynamics of farming systems were the most accurate and useful. Mapping of irrigated land use can be undertaken with classification of satellite images, though these methods may overlook small fields when applied with moderate to low resolution sensors. Smallholder irrigated areas were identified using composite images and high-dimensional statistics of temporal variation. This produced irrigated area maps with overall accuracy of 95.9% for test sites in Zimbabwe. The spectral median absolute deviation, as a measure of temporal variation on irrigated lands, was an important predictor of irrigated area. Incorporating within-season temporal variation into a machine learning prediction for light use efficiency also improved the accuracy of Gross Primary Productivity estimation for croplands.

Présentation

Version intégrale ( 40 Mb)