Designing novel classification approaches for mapping smallholder farms
Reliable crop maps are crucial for informed decision-making and to support sustainable farming. This, in turn, contributes to improving smallholder farmer livelihoods and the welfare of their families.
Such crop maps are typically developed by classifying remotely sensed images. However, conventional classification methods do not provide the accuracy required for large-scale mapping of smallholder farms in Africa. Complex crop-planting patterns, mixed crops, small fields, and vaguely delineated fields provide additional challenges to that of mapping crops from space. Two MSc students are currently working on novel classification approaches for mapping smallholder farms.
Figure 1. Example of mixed cropping in Sukumba community, Mali (source: ICRISAT)
Approach 1: designing a multi-classifier system
The objective of Rosa Aguilar is to design a multi-classifier system that combines several machine learning classifiers to, hopefully, produce a more accurate crop map than any map produced by a single classifier. This multi-classifier system is implemented on a cloud platform, which makes it possible to handle large sets of spatial and spectral features. A time series of very high spatial resolution images (WorldView) obtained over the Mali field site is used to illustrate the work. Rosa’s current experiments try find out the most optimal configuration of the training strategy, of the feature space, and of the rules to combine the individual classifiers.
Approach 2: investigating one class classifiers
In some situations, we are only interested in mapping a single crop, either because it is the most important crop in the area or because we do not have enough training samples to map multiple crops. In such cases, one-class classifiers can be used.
The objective of Kanmani Balasubramanian is to investigate the use of one-class (maize) classifiers in our Mali site. In addition, her work aims to identify the most important spectral and textural features for this task. This is because spectral (reflectance-based) and textural (pattern-based) features may possess unique and distinct characteristics that can help to discriminate between specific crops. With focus on important features, she hopes to reduce the complexity of the classification process.
Rosa and Kanmani are MSc students at the University of Twente, Faculty of Geoinformation Science and Earth Observation (ITC) and their research work is being co-supervised by Raul Zurita-Milla, Emma Izquierdo Verdiguier and Rolf de By. After they have finished their thesis, the results of Rosa’s and Kanmani’s work will be published in the STARS Knowledge Portal. You can find summaries of other student’s work in this section of the Knowledge Portal as well. If you are interested in (one of) these student’s work and want to know more, please do not hesitate to contact us!