A major challenge in automatic irrigation of extensive agricultural fields is large scale soil moisture monitoring. Proper orthogonal decomposition (POD) is a widespread data-driven dimension reduction technique which can be combined with QR pivoting method for estimation of high-dimensional signals and optimal sensor placement. However, it requires computation of tailored basis functions which should be extracted from known training data. This is feasible for problems with constant features such as face recognition. However, using fixed bases (and probably fixed sensor selection) may not be an appropriate approach for estimation of signals with time-variant features. This paper demonstrates that the POD algorithm can be implemented adaptively in a reliable fashion to address this issue. The probability of the mean estimation error to remain within a predetermined threshold is studied and theoretical results on permitted update intervals are provided. The adaptive tailored basis and sensor selection increases suitability to large scale applications with time-variant features. The performance of the algorithm on a NASA soil moisture dataset is analyzed.
- Adaptive sensor selection
- Proper orthogonal decomposition
- Reliable monitoring
- Soil moisture estimation
- Sparse monitoring