Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Integration of multi-source data with machine learning for agricultural drought monitoring in Morocco

Wang, Yuzhu LU (2025) In Student thesis series INES NGEM01 20251
Dept of Physical Geography and Ecosystem Science
Abstract
Drought is one of the most costly disasters worldwide. The complexity of nonlinear
relationships between drought variables and drought severity during drought events
poses significant challenges for accurate drought monitoring and prediction. Drought
research in Morocco has been dominated by linear and conventional statistical
methods, such as the pearson correlation. Although some earlier research has adopted
machine learning algorithms, there remains a lack of region-specific quantification
across different agricultural zones. This study aims to address these challenges by
using machine learning approaches to construct and validate an approximate model
and verification framework based on physiological hypotheses. I propose four
... (More)
Drought is one of the most costly disasters worldwide. The complexity of nonlinear
relationships between drought variables and drought severity during drought events
poses significant challenges for accurate drought monitoring and prediction. Drought
research in Morocco has been dominated by linear and conventional statistical
methods, such as the pearson correlation. Although some earlier research has adopted
machine learning algorithms, there remains a lack of region-specific quantification
across different agricultural zones. This study aims to address these challenges by
using machine learning approaches to construct and validate an approximate model
and verification framework based on physiological hypotheses. I propose four
machine learning-based models, with a linear regression model, to comprehensively
evaluate model performance across different growth stages of winter wheat and assess
regional variations in model effectiveness. Specifically, Monthly remote sensing
indices—Vegetation Condition Index (VCI), Gross Primary Productivity Condition
Index (GPPCI), Soil Moisture Condition Index (SMCI), and Temperature Condition
Index (TCI)—are used to reflect crop responses to drought across different growth
stages. All models demonstrated the capability to predict drought at least three months
in advance. At the scale of Morocco's agricultural regions, feature importance analysis
revealed that the TCI played a dominant role in the early growth stages, while the VCI
became more influential in the later stages. Among all variables, GPPCI in March
consistently showed the highest explanatory power throughout the growing season. Furthermore, the models were able to capture temporal lags and regional differences
in winter wheat responses to drought. For example, in mountainous regions, soil
moisture emerged as a key factor during the early growth stages, whereas in the
southern marginal zones, early-season temperature stress had a more pronounced
impact on crop development. (Less)
Popular Abstract
Every year, Moroccan farmers face uncertainty as rainfall fluctuates and drought risks rise. Wheat, the country’s staple crop, is especially vulnerable. To help monitor and manage these risks, this study uses satellite data and machine learning to develop a new tool for identifying drought stress early and accurately.
Remote sensing satellites collect data on how vegetation grows, how warm the land surface is, and how much moisture is in the soil. By combining these data sources, the study trained computer models to recognize drought patterns and predict how much crops might suffer. The models look at the entire growing season—from planting in November to harvesting in June—and learn which months and variables are most important for... (More)
Every year, Moroccan farmers face uncertainty as rainfall fluctuates and drought risks rise. Wheat, the country’s staple crop, is especially vulnerable. To help monitor and manage these risks, this study uses satellite data and machine learning to develop a new tool for identifying drought stress early and accurately.
Remote sensing satellites collect data on how vegetation grows, how warm the land surface is, and how much moisture is in the soil. By combining these data sources, the study trained computer models to recognize drought patterns and predict how much crops might suffer. The models look at the entire growing season—from planting in November to harvesting in June—and learn which months and variables are most important for drought prediction.
Interestingly, the study found that March is a key turning point: temperature and soil moisture dominate in the early season, but in March, plant productivity becomes the best indicator of drought. By April and May, visible signs of stress like poor leaf growth are easier to detect. The models also revealed that different regions respond differently to drought. For example, mountains rely more on shallow soil moisture, while southern zones are sensitive to heat.
The final product is a flexible drought monitoring index that doesn’t rely on weather station data. This means it can be used in remote or data-scarce regions. With better early warnings, farmers and decision-makers can better prepare for tough growing seasons—and improve resilience to climate change. (Less)
Please use this url to cite or link to this publication:
author
Wang, Yuzhu LU
supervisor
organization
course
NGEM01 20251
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Physical Geography and Ecosystem Analysis, agricultural drought prediction, machine learning, new drought index, remote sensing.
publication/series
Student thesis series INES
report number
719
language
English
id
9203817
date added to LUP
2025-06-24 15:41:44
date last changed
2025-06-24 15:41:44
@misc{9203817,
  abstract     = {{Drought is one of the most costly disasters worldwide. The complexity of nonlinear
relationships between drought variables and drought severity during drought events
poses significant challenges for accurate drought monitoring and prediction. Drought
research in Morocco has been dominated by linear and conventional statistical
methods, such as the pearson correlation. Although some earlier research has adopted
machine learning algorithms, there remains a lack of region-specific quantification
across different agricultural zones. This study aims to address these challenges by
using machine learning approaches to construct and validate an approximate model
and verification framework based on physiological hypotheses. I propose four
machine learning-based models, with a linear regression model, to comprehensively
evaluate model performance across different growth stages of winter wheat and assess
regional variations in model effectiveness. Specifically, Monthly remote sensing
indices—Vegetation Condition Index (VCI), Gross Primary Productivity Condition
Index (GPPCI), Soil Moisture Condition Index (SMCI), and Temperature Condition
Index (TCI)—are used to reflect crop responses to drought across different growth
stages. All models demonstrated the capability to predict drought at least three months
in advance. At the scale of Morocco's agricultural regions, feature importance analysis
revealed that the TCI played a dominant role in the early growth stages, while the VCI
became more influential in the later stages. Among all variables, GPPCI in March
consistently showed the highest explanatory power throughout the growing season. Furthermore, the models were able to capture temporal lags and regional differences
in winter wheat responses to drought. For example, in mountainous regions, soil
moisture emerged as a key factor during the early growth stages, whereas in the
southern marginal zones, early-season temperature stress had a more pronounced
impact on crop development.}},
  author       = {{Wang, Yuzhu}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Student thesis series INES}},
  title        = {{Integration of multi-source data with machine learning for agricultural drought monitoring in Morocco}},
  year         = {{2025}},
}