Pedro Machado

Image-based plant-disease classifiers often achieve high accuracy under controlled conditions, but their reliability can deteriorate in real-world field environments. This article investigates crossdataset strawberry leaf-health classification using controlled PlantVillage images and field images collected through the SmartBerry project in Nigeria. Four deep-learning architectures were evaluated under within-domain, controlled-to-field and pooled-training conditions using healthy-leaf and darkspot/ scorch symptoms. Although all models achieved near-ceiling performance on PlantVillage, direct transfer to SmartBerry reduced macro F1-score to between 0.418 and 0.824, demonstrating a substantial gap between controlled-dataset performance and field reliability. To reduce the amount of locally labelled field data required to address this gap, this paper proposes SmartBerry Domain Adaptation (SmartBerry-DA). This label-efficient adaptation approach combines a model trained on controlled images with a small set of labelled SmartBerry images and additional unlabelled field images. Using MobileViTv2-0.5, SmartBerry-DA achieved a macro F1-score of 0.940 ± 0.015 with only five labelled SmartBerry images per class and 0.988 ± 0.012 with ten, compared with 0.994±0.008 when the full labelled target dataset was used. This shows that strong field performance can be recovered with substantially reduced local annotation. For lightweight offline deployment, 16- bit floating-point representation reduced model size from 4.37 MB to 2.30 MB without changing predictive performance under the evaluated conditions. The findings demonstrate the importance of locally acquired field data and label-efficient adaptation for transferring agricultural AI from benchmark datasets to practical farming environments, while providing a pathway towards lightweight strawberry disease recognition in settings where extensive annotation and continuous connectivity may not be feasible.

Zoos and aquaria must retain genetic diversity for the species they house if the populations are to remain viable. The cryopreservation and storage of living animal cells is a potential strategy for minimising genetic diversity loss over time, when genetic diversity can be effectively reintroduced in the future as population genetic diversity decreases. Using population simulations, this study investigates the potential impact of using cryopreserved sperm on genetic diversity retention in 18 species of pupfish: a priority taxonomic group for ex‐situ conservation management for the European Association of Zoos and Aquaria (EAZA). First, genetic diversity simulations were conducted to determine the necessary population sizes required for each species to achieve a 90% genetic diversity target over a short (30‐year) or longer (100‐year) period without the integration of cryopreserved sperm into breeding management, with Ne/N simulated at either 0.05 (worst‐case scenario), 0.1 or 0.3 (best‐case scenario). Subsequently, these simulations were then compared against the required population sizes to achieve the same genetic diversity targets if cryopreserved sperm is used for 100%, 50%, 10% or 1% of breeding attempts. The simulations demonstrate that no species are currently being kept in sufficient numbers to achieve 90% genetic diversity. Specifically, the simulations projected that across the priority species ( n = 18) the average genetic diversity per species by 100 years would be 1.16% (±2.48% SD), 6.13% (±9.25% SD) and 23.57% (±25.28% SD) for Ne/N = 0.05, 0.1 and 0.3 respectively. To achieve the 90% target over 100 years, without the use of cryopreserved sperm (cryoARTs), the mean total number of fish required was 9498, 4747 and 1582 fish if Ne/N = 0.05, 0.1 or 0.3, respectively. Overall, the mean number of fish per species was 138.61 individuals (±128.61 SD) and to achieve the 90% genetic diversity target over 100 years, Ne/N = 0.3, all populations need to be increased by approximately 1400 individuals. Collectively, these scenarios demonstrate rapid loss of genetic diversity unless populations are substantially increased. Across the priority species ( n = 18), if cryoARTs is used at 10% of breeding attempts then the average total number of individual fish required to achieve the 90% genetic diversity target, for Ne/N = 0.05, 0.1 or 0.3, is 910, 455 and 152 fish respectively, over a 100‐year period. This study demonstrates that investment in greater pupfish population sizes in EAZA collections, alongside a strategy involving the integration of cryoARTs into species management, could aid in population viability by improving genetic diversity prospects.

Secure data obfuscation requires balancing perceptual transparency, computational efficiency, and architectural deployability. Conventional spatial-domain steganography achieves high capacity but lacks structural abstraction and hardware-oriented design. We introduce SteganoSNN, a neuromorphic steganographic framework that integrates spike-based temporal encoding with field programmable gate array (FPGA)-accelerated embedding for secure and energy-aware multimedia data hiding. Audio samples are transformed into spike-count regimes using a Leaky Integrate-and-Fire (LIF) spiking neuron model, followed by modulo-based symbol transformation and deterministic spike-index mapping. The spike-to-Spike Index (SI) abstraction decouples source symbols from embedded bit patterns, introducing a temporal encoding layer prior to least significant bit (LSB) embedding in red, green, blue, alpha (RGBA) images. The proposed system is implemented in Python using NEST and realised on a PYNQ-Z2 FPGA through a hybrid processing-system/programming-logic co-design. Post-implementation analysis reports total on-chip power consumption of approximately 1.41 W, with programmable logic contributing only a small fraction of overall energy usage. The proposed framework supports an embedding capacity of 8 bits per pixel (bpp) while maintaining high perceptual fidelity on the DIV2K 2017 dataset, achieving peak signal-to-noise ratio (PSNR) values between 40.4 dB and 41.35 dB and structural similarity index (SSIM) above 0.97. Real-time feasibility is demonstrated through detailed latency and throughput evaluation on full-High Definition (HD) images. Unlike purely LSB-based approaches, SteganoSNN introduces a neuromorphic temporal representation layer that enables hardware-efficient symbol abstraction without increasing embedding depth. The results establish spike-based encoding as a viable architectural paradigm for secure and resource-aware steganography in edge-Artificial Intelligence (AI) and embedded systems.

Monitoring behavioural drift, a sustained shift in an individual’s daily activity, sleep, or social patterns, offers a significant lens for early mental health intervention. However, detecting these drifts in free-living settings remains challenging due to the absence of ground-truth labels, the temporal complexity of human behaviour, and fragmentation across heterogeneous sensing modalities. This paper proposes a multimodal approach to quantify and detect behavioural drift using longitudinal data from over 500 university students in the NetHealth cohort. We extract personalised, longitudinal features spanning three behavioural domains: physical activity, sleep hygiene, and communication diversity and model deviations relative to rolling, individual-specific statistical baselines. To differentiate transient anomalies from meaningful behavioural change, we introduce a sustained streak mechanism that identifies persistent drift episodes. We evaluate three temporal modelling strategies: Isolation Forest, Convolutional Neural Networks, and Long Short-Term Memory networks across both single-modality and fused approaches. Our findings indicate that recurrent models offer the strongest performance, highlighting the necessity of capturing temporal dependencies in behavioural data. Furthermore, we find that cross-modal correlations between drift signals are weak, confirming that activities, sleep, and communication provide complementary, non-redundant insights into an individual’s well-being. This work establishes the methodological basis for integrating multimodal sensing data to monitor mental health trajectories, providing a scalable path towards early intervention in digital health.

With increasing life expectancy, particularly in developed nations, the proportion of elderly individuals is rising rapidly, necessitating advanced systems for continuous monitoring and timely intervention to support independent living and enhance safety in assisted care environments. Falls are among the leading causes of hospitalizations and deaths related to injuries in this demographic, highlighting the urgent need for intelligent fall detection systems. However, most existing solutions struggle with real-world deployment due to incomplete anomaly modeling and a lack of contextual location awareness. This article introduces a novel position-aware indoor activity recognition and fall detection approach that uses spatial and motion data to detect falls with high accuracy and contextual relevance. The system integrates ultrawideband (UWB) positioning technology with a multilayer perceptron (MLP) model to achieve indoor localization. Furthermore, accelerometer and gyroscope data are used for activity monitoring, which is processed using a hybrid deep learning (DL) architecture that combines a variational autoencoder (VAE), convolutional neural networks (CNNs), and long short-term memory (LSTM) networks. This architecture takes advantage of temporal and spatial feature extraction for improved fall detection. The localization module achieves over 96% accuracy. For activity recognition, the VAE CNN-LSTM model achieves fall detection accuracy exceeding 97%. A late fusion decision layer combines spatial and activity-level insights to enable precise detection and localization of fall events within indoor environments. The proposed system is validated in a real-world smart home setting and demonstrates strong performance in terms of accuracy, scalability, and adaptability.

Monitoring behavioural drift, the sustained shift in an individual’s daily activity, sleep, or social patterns offers a significant lens for early mental health intervention. However, detecting these drifts in free-living settings remains challenging due to the absence of ground-truth labels, the temporal complexity of human behaviour, and fragmentation across heterogeneous sensing modalities. This paper proposes a multimodal approach to quantify and detect behavioural drift using longitudinal data from over 500 university students in the NetHealth cohort. We extract personalised, longitudinal features spanning three behavioural domains physical activity, sleep hygiene, and communication diversity and model deviations relative to rolling, individual-specific statistical baselines. To differentiate transient anomalies from meaningful behavioural change, we introduce a sustained streak mechanism that identifies persistent drift episodes. We evaluate three temporal modelling strategies Isolation Forest, Convolutional Neural Networks, and Long Short-Term Memory networks across both single-modality and fused approaches. Our findings indicate that recurrent models offer the strongest performance, highlighting the necessity of capturing temporal dependencies in behavioural data. Furthermore, we find that cross-modal correlations between drift signals are weak, confirming that activity, sleep, and communication provide complementary, non-redundant insights into an individual's wellbeing. This work establishes a robust methodological basis for integrating multimodal sensing data to monitor mental health trajectories, providing a scalable path toward early intervention in digital health.

Alzheimer’s Disease (AD) represents a growing global health challenge, driven by complex genetic factors and diverse risk contributors. Currently, an estimated 55 million people worldwide are affected by dementia, with AD responsible for 60–70% of these cases. This paper explores the application of advanced machine learning approaches to predict AD risk using Genome-Wide Association Studies data from multiple cohorts, with a particular focus on transfer learning and feature selection techniques. We evaluate the performance of Wide and Deep Neural Networks and Multi-Head Attention in assessing their ability to generalise across datasets. As part of this, we explore knowledge distillation as a strategy to enhance model efficiency through improved generalisation performance in smaller architectures by transferring knowledge from high-capacity models to lightweight ones. Furthermore, the performance of these deep learning approaches is compared with tree-based ensembles, including Random Forest and XGBoost. Our experiments evaluate the generalisability, transferability, and efficiency of these models across different transfer learning scenarios. Findings indicate that aggregating multi-cohort training data significantly enhances predictive performance, highlighting the importance of data diversity in improving AD risk assessment. The proposed knowledge distillation approach enables the transfer of knowledge from a complex teacher model to a simpler student model, significantly improving performance. To enhance interpretability, we apply SHAP (SHapley Additive exPlanations) to the student models, revealing cohort-specific differences in SNP importance and highlighting variants in genes such as ABI3BP and SYN3, both of which are linked to immune and synaptic functions in AD. The integration of SHAP enables transparent interpretation of model decisions and supports the identification of transferable genetic markers, reinforcing the clinical relevance of our framework in AD risk prediction.