UAV multispectral mapping and field-level photography for crop-type and intercropping detection supporting agricultural statistics in Mozambique
The topic is also suitable for GEM students in track 3 – GEM for Ecosystems & Natural Resources.
Smallholder farming systems in Sub-Saharan Africa frequently involve intercropping, where multiple crops are grown together on the same parcel. These diverse systems enhance resilience and biodiversity and are fundamental to local food security, yet they pose major challenges for both agricultural statistics and remote sensing. Satellite images often contain mixed pixels at very small scales, making it difficult to separate crop types, while field-level surveys are costly and very dependent on training of the field enumerators.
The Statistics from Space project seeks to address these challenges by integrating UAV multispectral imagery, ground-level photography and structured data collected by field enumerators to build high-quality reference datasets for agricultural monitoring.
In Manica and Gaza provinces (Mozambique), enumerators recorded crop types such as maize, millet, beans, cassava and groundnut and documented plot conditions, while also capturing thousands of field photographs with smartphones.
Parallel to this, extensive UAV campaigns were conducted using a DJI Mavic 3 Multispectral drone. These flights produced high-resolution RGB and multispectral (Red Edge, NIR, Red and Green) orthomosaics, together with digital surface models and point clouds, all at a pixel size of approximately five centimetres. This level of detail makes it possible to observe fine-scale crop arrangements and the spatial structure of intercropped systems.
The MSc project aims to transform this multi-sensor dataset into improved agricultural statistics for intercropping environments. One key direction is the development of semantic segmentation models, such as U-Nets or similar architectures, to detect individual crop types within complex mixtures. Building on this, the student may explore multi-label classification approaches to capture situations where more than one species is present in a single pixel. Another focus lies in modelling fractional cover to quantify the relative area of each crop type. For example, estimating the proportions of maize and groundnut within the same plot. Field photographs can serve as an additional source of information, either by validating spectral patterns or by supporting cross-view learning approaches that link ground perspective to overhead imagery. The resulting crop maps can then be translated into area estimates that feed directly into agricultural statistics and serve as training data for later upscaling to satellite platforms.
Finally, the project offers space to examine how well models generalize across regions, for instance by comparing performance between Manica and Gaza provinces.
This MSc research contributes to the modernisation of agricultural statistics in Africa by developing reproducible, scalable methods that integrate UAV data, field observations and machine learning to improve the understanding of complex smallholder cropping systems.
The main objective of this research is to develop and evaluate image-based models capable of detecting crop types and intercropped field compositions in smallholder agricultural systems in Mozambique. The topic is part of the Statistics from Space project and builds on a rich existing dataset: thousands of labelled field photographs, field-enumerator crop labels and processed UAV multispectral orthomosaics. A real treasure of agricultural data!
The student will explore deep learning approaches, particularly U-Net or similar segmentation architectures to:
- Detect crop types in UAV multispectral imagery (e.g., maize, millet, cassava, beans, groundnut).
- Identify crop compositions in intercropped systems (e.g., maize/groundnut mixtures) using both UAV imagery and field photographs.
- Estimate fractional cover (FC) for each crop type, enabling calculation of crop area proportions within mixed plots.
- Translate these FC estimates into crop statistics, such as crop area distributions and smallholder cropping patterns.
- Prepare proxy training datasets that can later be used to upscale insights to Sentinel-2 or other satellite platforms.
The methodology includes:
- Integrating UAV orthomosaics, field photos and enumerator labels.
- Training segmentation, classification or multi-label models for mixed cropping environments.
- Validating model outputs against field measurements and enumerator records.
- Producing crop-area estimates and assessing how intercropping can be best included into agricultural statistics.
- (Optional) Testing domain adaptation approaches to prepare training data for satellite sensors.
All UAV data (RGB, Red Edge, NIR, Green, Red) and processing products (DSM, orthomosaics, point clouds) are already available, saving considerable time and allowing the student to focus on modelling and analysis.