In this article, we explore the role and usefulness of neuro-fuzzy logic in the context of automatically reasoning under uncertainty about complex scenes in remotely sensed data. Specifically, we consider a first order Takagi- Sugeno-Kang (TSK) adaptive neuro-fuzzy inference system (ANFIS). First, we explore the idea of embedding an experts knowledge into ANFIS. Second, we explore the augmentation of this knowledge via optimization relative to training data. The aim is to explore the possibility of transferring then improving domain performance on tedious but important and challenging tasks. This route was selected, versus the popular modern thinking of learning a neural solution from scratch in an attempt to maintain interpretability and explainability of the resultant solution. An additional objective is to observe if the machine learns anything that can be returned to the human to improve their individual performance. To this end, we explore the task of detecting construction sites, an abstract concept that has a large amount of inner class variation. Our experiments show the usefulness of the proposed methodology and it sheds light onto future directions for neuro-fuzzy computing, both with respect to performance, but also with respect to glass box solutions.
Developing accurate methods to determine bathymetry, bottom type, and water column optical properties from hyperspectral imagery is an ongoing scientific problem. Recent advances in deep learning have made convolutional neural networks (CNNs) a popular method for classification and regression on complex datasets. In this paper, we explore the use of CNNs to extract water depth, bottom type, and inherent optical properties (IOPs) from hyperspectral imagery (HSI) of water. We compare the CNN results to other machine learning algorithms: k-nearest-neighbors (KNN), stochastic gradient descent (SGD), random forests (RF), and extremely randomized trees (ET). This work is an inverse problem in which we seek to find the water properties than impact the reflectance and hence the collected HSI. The data includes optically shallow water, in which the bottom can be seen, and optically deep, in which the bottom cannot be seen and does not affect the reflectance. The scalar optical properties we find through regression are chlorophyll (CHL), colored dissolved organic matter (CDOM), total suspended sediments (TSS). For the case of the optically shallow water, we classify the bottom type among 114 different substrates. The results demonstrate that for finding water depth, bottom type, and IOPs in the case of optically shallow water, the CNN has better performance than other machine learning methods. For regression of the IOPs in optically deep water, the extremely randomized trees method outperforms the CNN. We further investigate the mechanisms of these results and discuss hyperparameter tuning strategies that may improve deep learning accuracy.
Many aspects of decision making processes for autonomous systems involve human subjective information in some form. Methods for informing decision making processes with human information are needed to inform probabilistic information used in an autonomous system. This can provide better decisions and permit a UAV to more quickly and efficiently complete tasks. Specifically we use possibility theory to represent the subjective information and apply possibilistic conditioning of the probability distribution. A simulation platform was developed to evaluate approaches to using possibilistic inputs and showed that is was feasible to make effective usage of such information.
Deep Learning has proven to be an effective method for making highly accurate predictions from complex data sources. Convolutional neural networks continue to dominate image classification problems and recursive neural networks have proven their utility in caption generation and language translations. While these approaches are powerful, they do not offer explanation for how the output is generated. Without understanding how deep learning arrives at a solution there is no guarantee that these networks will transition from controlled laboratory environments to fieldable systems. This paper presents an approach for incorporating such rule based methodology into neural networks by embedding fuzzy inference systems into deep learning networks.