Optimizing supervised quantum machine learning for pixel classification

Mapping products for flight simulators require feature classification for LandUse/LandCover (LULC) and sensor simulation. The challenge lies in providing accurate material classification of imagery, while significantly reducing the analyst processing time and cost. Future quantum computers may provide sufficient speed for processing every pixel simultaneously. Image segmentation can be thought of as pixel clustering with each pixel label corresponding to a particular LULC class. We characterize a supervised quantum machine learning (QML) use case with imagery and lidar. The fast and accurate segmentation of an image into just a few LULC classes (e.g., buildings, bare earth, roads, water, vegetation) can benefit material classification applications. We classically simulate a binary classification quantum circuit as part of a quantum deep learning layer to help distinguish LULC features. We also use an ensemble of quantum neural network (QNN) machine-learning models with different solvers. We then use a linear program (LP) approximation circuit to optimally decide which model to use per pixel in a one-sided-versus-nature game-theoretic implementation. This input matrix is generated from goodness-of-fitness probability metrics. Furthermore, we perform multiclass feature decision making for our quantum algorithm using a binary comparison matrix. We process a pairwise reward matrix of prediction probability confidences to choose the pixel class. The optimal decision making from these matrices uses an LP approximation subset-summing method with quantum adders and comparators. As the size of the reward matrix increases, LP approximation approaches the performance of a classical LP. We compare classical and quantum image segmentation accuracy results with confusion matrices.