Multi-scale crop stress detection using soil moisture networks, weather stations, satellite data and optional UAV imaging
The topic is also suitable for GEM students in track 3 – GEM for Ecosystems & Natural Resources.
Understanding how crops and grasslands respond to drought and weather extremes requires observations at multiple scales. In the Twente region, a new environmental monitoring network (ITEM, operational from early 2026) will provide continuous meteorological measurements, while existing IoT soil moisture sensors already offer a longer time series of subsurface conditions. When combined with satellite thermal and multispectral observations, these ground-based measurements allow us to detect early signs of crop or meadow stress before major yield or biomass losses occur.
This data set may optionally be complemented with small UAV campaigns or repeat photography from simple camera traps. This additional data could provide higher-resolution views of canopy conditions, but the thesis remains fully feasible without UAV flights.
The central idea is to fuse ground-based sensor data with remote sensing indicators and apply explainable machine learning to detect or predict drought impacts, weather-driven stress or growth anomalies. Because the ITEM network is new, the project will rely primarily on soil moisture time series and satellite data, enriched with the first available meteo observations and optional student-collected imagery. The thesis will focus on one or a small number of representative test sites in the Twente region.
This is a rapidly growing research area with excellent career opportunities in precision agriculture, environmental modelling and AI-driven decision support systems.
The objectives of the topic are as follows:
(1) Assemble a multi-sensor dataset by integrating historical soil moisture data, newly available meteorological measurements from the ITEM network and selected satellite indices such as NDVI, NDWI, or Land Surface Temperature. If desired, complementary UAV imagery or repeat photographs from simple camera traps can also be collected to obtain higher-resolution canopy information.
(2) Derive environmental stress indicators from each data source, including soil moisture anomalies, vegetation indices, canopy temperature metrics and basic phenological or greenness measures from repeat photography where available.
(3) Develop and train an explainable machine learning model to detect or predict crop or meadow stress. The model should use interpretable approaches such as feature importance, SHAP values, or rule-based methods to provide transparent insights.
(4) Evaluate the model at one or a small number of field sites, assessing how well the integration of multi-scale sensing improves early stress detection and quantifying the added value of optional high-resolution imagery compared to using only ground sensors and satellite data.