Michal Myller

Monitoring crop irrigation levels and the soil moisture in the root zone is of paramount importance in precision agriculture (PA), as it allows practitioners to optimize the water supply. This can, in turn, directly lead to significant water savings while maintaining appropriate cultivation practices. Currently adopted in-field methods to monitor root-zone soil moisture are costly, human-dependent, and unscalable to large spatial areas. We tackle this issue and propose MoleNose---an approach for estimating root-zone soil moisture from hyperspectral images using deep learning models enhanced through particle swarm optimization. Our experimental study, performed over a real-world dataset of hyperspectral images collocated with in-situ soil moisture measurements, revealed that MoleNose significantly outperforms hand-crafted deep learning architectures in various temporal analysis scenarios, in which we estimate the root-zone soil moisture in different time points in the season.

Enhancing agricultural methods through the utilization of Earth observation and artificial intelligence (AI) has emerged as a significant concern. The ability to quantify soil parameters on a large scale can play a pivotal role in optimizing the fertilization process. While techniques for noninvasive estimation of soil parameters from hyperspectral images (HSIs) exist, their validation typically occurs across different datasets and employs varying validation protocols. This diversity renders them inherently challenging (or even impossible) to compare objectively.

Hyperspectral images capture very detailed information about scanned objects and, hence, can be used to uncover various characteristics of the materials present in the analyzed scene. However, such image data are difficult to transfer due to their large volume, and generating new ground-truth datasets that could be utilized to train supervised learners is costly, time-consuming, very user-dependent, and often infeasible in practice. The research efforts have been focusing on developing algorithms for hyperspectral data classification and unmixing, which are two main tasks in the analysis chain of such imagery. Although in both of them, the deep learning techniques have bloomed as an extremely effective tool, designing the deep models that generalize well over the unseen data is a serious practical challenge in emerging applications. In this paper, we introduce the deep ensembles benefiting from different architectural advances of convolutional base models and suggest a new approach towards aggregating the outputs of base learners using a supervised fuser. Furthermore, we propose a model augmentation technique that allows us to synthesize new deep networks based on the original one by injecting Gaussian noise into the model’s weights. The experiments, performed for both hyperspectral data classification and unmixing, show that our deep ensembles outperform base spectral and spectral-spatial deep models and classical ensembles employing voting and averaging as a fusing scheme in both hyperspectral image analysis tasks.

Benchmarking deep learning algorithms before deploying them in hardware-constrained execution environments, such as imaging satellites, is pivotal in real-life applications. Although a thorough and consistent benchmarking procedure can allow us to estimate the expected operational abilities of the underlying deep model, this topic remains under-researched. This paper tackles this issue and presents an end-to-end benchmarking approach for quantifying the abilities of deep learning algorithms in virtually any kind of on-board space applications. The experimental validation, performed over several state-of-the-art deep models and benchmark datasets, showed that different deep learning techniques may be effectively benchmarked using the standardized approach, which delivers quantifiable performance measures and is highly configurable. We believe that such benchmarking is crucial in delivering ready-to-use on-board artificial intelligence in emerging space applications and should become a standard tool in the deployment chain.

Although hyperspectral images capture very detailed information about the scanned objects, their efficient analysis, transfer, and storage are still important practical challenges due to their large volume. Classifying and segmenting such imagery are the pivotal steps in virtually all applications, hence developing new techniques for these tasks is a vital research area. Here, deep learning has established the current state of the art. However, deploying large-capacity deep models on-board an Earth observation satellite poses additional technological challenges concerned with their memory footprints, energy consumption requirements, and robustness against varying-quality image data, with the last problem being under-researched. In this paper, we tackle this issue, and propose a set of simulation scenarios that reflect a range of atmospheric conditions and noise contamination that may ultimately happen on-board an imaging satellite. We verify their impact on the generalization capabilities of spectral and spectral-spatial convolutional neural networks for hyperspectral image segmentation. Our experimental analysis, coupled with various visualizations, sheds more light on the robustness of the deep models and indicate that specific noise distributions can significantly deteriorate their performance. Additionally, we show that simulating atmospheric conditions is key to obtaining the learners that generalize well over image data acquired in different imaging settings.

Hyperspectral image analysis has become an important topic widely researched by the remote sensing community. Classification and segmentation of such imagery help understand the underlying materials within a scanned scene since hyperspectral images convey detailed information captured in a number of spectral bands. Although deep learning has established the state-of-the-art in the field, it still remains challenging to train well-generalizing models due to the lack of ground-truth data. In this letter, we tackle this problem and propose an end-to-end approach to segment hyperspectral images in a fully unsupervised way. We introduce a new deep architecture which couples 3-D convolutional autoencoders with clustering. Our multifaceted experimental study--performed over the benchmark and real-life data--revealed that our approach delivers high-quality segmentation without any prior class labels.

Data augmentation helps improve generalization capabilities of deep neural networks when only limited ground-truth training data are available. In this letter, we propose test-time augmentation of hyperspectral data, which is executed during the inference rather than before the training of deep networks. We introduce two augmentation techniques, which can be applied at both training time and test time. The experiments revealed that our augmentations boost generalization of deep models and work in real time, and the test-time approach can be combined with training-time techniques to enhance the classification accuracy.