Search for: All records

Accurate uncertainty quantification is necessary to enhance the reliability of deep learning (DL) models in realworld applications. In the case of regression tasks, prediction intervals (PIs) should be provided along with the deterministic predictions of DL models. Such PIs are useful or “high-quality (HQ)” as long as they are sufficiently narrow and capture most of the probability density. In this article, we present a method to learn PIs for regression-based neural networks (NNs) automatically in addition to the conventional target predictions. In particular, we train two companion NNs: one that uses one output, the target estimate, and another that uses two outputs, the upper and lower bounds of the corresponding PI. Our main contribution is the design of a novel loss function for the PI-generation network that takes into account the output of the target-estimation network and has two optimization objectives: minimizing the mean PI width and ensuring the PI integrity using constraints that maximize the PI probability coverage implicitly. Furthermore, we introduce a self-adaptive coefficient that balances both objectives within the loss function, which alleviates the task of fine-tuning. Experiments using a synthetic dataset, eight benchmark datasets, and a real-world crop yield prediction dataset showed that our method was able to maintain a nominal probability coverage and produce significantly narrower PIs without detriment to its target estimation accuracy when compared to those PIs generated by three state-of-the-art neuralnetwork-based methods. In other words, our method was shown to produce higher quality PIs. 

Hyperspectral imaging systems are becoming widely used due to their increasing accessibility and their ability to provide detailed spectral responses based on hundreds of spectral bands. However, the resulting hyperspectral images (HSIs) come at the cost of increased storage requirements, increased computational time to process, and highly redundant data. Thus, dimensionality reduction techniques are necessary to decrease the number of spectral bands while retaining the most useful information. Our contribution is two-fold: First, we propose a filter-based method called interband redundancy analysis (IBRA) based on a collinearity analysis between a band and its neighbors. This analysis helps to remove redundant bands and dramatically reduces the search space. Second, we apply a wrapper-based approach called greedy spectral selection (GSS) to the results of IBRA to select bands based on their information entropy values and train a compact convolutional neural network to evaluate the performance of the current selection. We also propose a feature extraction framework that consists of two main steps: first, it reduces the total number of bands using IBRA; then, it can use any feature extraction method to obtain the desired number of feature channels. We present classification results obtained from our methods and compare them to other dimensionality reduction methods on three hyperspectral image datasets. Additionally, we used the original hyperspectral data cube to simulate the process of using actual filters in a multispectral imager. 

Optical sensing has the potential to be an important tool in the automated monitoring of food quality. Specifically, hyperspectral imaging has enjoyed success in a variety of tasks ranging from plant species classification to ripeness evaluation in produce. Although effective, hyperspectral imaging is prohibitively expensive to deploy at scale in a retail setting. With this in mind, we develop a method to assist in designing a low-cost multispectral imager for produce monitoring by using a genetic algorithm (GA) that simultaneously selects a subset of informative wavelengths and identifies effective filter bandwidths for such an imager. Instead of selecting the single fittest member of the final population as our solution, we fit a univariate Gaussian mixture model to the histogram of the overall GA population, selecting the wavelengths associated with the peaks of the distributions as our solution. By evaluating the entire population, rather than a single solution, we are also able to specify filter bandwidths by calculating the standard deviations of the Gaussian distributions and computing the full-width at half-maximum values. In our experiments, we find that this novel histogram-based method for feature selection is effective when compared to both the standard GA and partial least squares discriminant analysis. 

While a great deal of research has been directed towards developing neural network architectures for RGB images, there is a relative dearth of research directed towards developing neural network architectures specifically for multi-spectral and hyper-spectral imagery. We have adapted recent developments in small efficient convolutional neural networks (CNNs), to create a small CNN architecture capable of being trained from scratch to classify 10 band multi-spectral images, using much fewer parameters than popular deep architectures, such as the ResNet or DenseNet architectures. We show that this network provides higher classification accuracy and greater sample efficiency than the same network using RGB images. Further, using a Bayesian version of our CNN architecture we show that a network that is capable of working with multi-spectral imagery significantly reduces the uncertainty associated with class predictions compared to using RGB images.