Performance Measure-Based Band Group Selection for Agricultural Multispectral Sensor Design

Matthew Allen Lee


Hyperspectral sensors are unfortunately plagued by relative high financial cost and they require large amounts of storage and computation. Too often, these factors limit their applicability in precision agriculture and in particular application requiring real-time signal processing. On the other hand, multispectral sensors are less expensive and require far less computational resources. If the task of the sensor or its platform is well known ahead of time, then it can be beneficial to design a multispectral sensor to accomplish the task as long as the performance is not degraded. In this article, we explore the design of a task specific multispectral sensor based on a new band grouping technique. Band grouping algorithms typically rely on a proximity measure to determine how similar (or dissimilar) the information contained in hyperspectral bands are to each other. Similar bands are then grouped. However, the proximity measure typically does not take into account the interactions of different band groups or how band groups will be used. The theory put forth in this article is unique because it makes global decisions for band groups by utilizing a performance measure to gauge the effectiveness of random partitionings at some given task, such as classification or anomaly detection. The band groups that are most correlated with good performance are then selected. Our technique is compared to the uniform partitioning technique using the Pecan1 data set. The results show that the overall accuracy is better with our technique when the same number  of band groups are selected.


Hyperspectral, band grouping, dimensionality reduction, herbicides

Full Text:



Indexadores / Base de Dados:


Google Scholar


Journal of Hyperspectral Remote Sensing - eISSN: 2237-2202