Projects

A selection of projects spanning imaging hardware, image restoration, segmentation and forecasting.

Low-Cost Dual-Spectrum Sky Imager

Most sky-imaging platforms capture a single spectral modality, such as visible light or infrared, and many are expensive. We designed a low-cost data acquisition system that records synchronised visible (RGB) images, near-infrared (NIR) images and environmental data every minute. The platform combines an RGB camera and an NIR imaging module running on Raspberry Pi 4B units with light, temperature and humidity sensors, all housed in a weatherproof enclosure for outdoor operation. Fusing models trained on RGB and NIR imagery reduced irradiance-estimation error by about 30% compared with RGB alone.

Paper: Measurement: Energy (2026), DOI · Code: dual-spectrum-solar-prediction

The dual-spectrum sky imager deployed outdoors: a weatherproof enclosure with two dome-covered cameras (RGB and near-infrared).

Architecture of SEE-U-Net: an encoder-decoder network built from Squeeze-Expand-Excite (SEE) blocks with skip connections, turning a sky image into a cloud mask.

Efficient Day & Night Cloud Segmentation

Cloud segmentation from ground-based sky images is essential for weather forecasting, solar energy prediction and atmospheric monitoring. However, many high-performing models are too computationally expensive for edge deployment, and night-time imaging adds low illumination and sensor noise. We proposed SEE-U-Net, an efficient encoder–decoder network built around a Squeeze-Expand-Excite (SEE) block. Compared with a standard U-Net, it uses 96.7% fewer parameters and over 45× fewer FLOPs while achieving higher accuracy on the SWINySEG day–night benchmark. We also introduced the Avro Nighttime Cloud Segmentation Dataset, a collection of night-time sky images from Ireland, Bangladesh and Australia for evaluating cross-region performance.

Earlier, we developed Tiny-UCloudNet, an ultra-lightweight U-Net variant built with depthwise separable convolutions, with a memory footprint of only about 7 KB of trainable parameters.

Papers: The Visual Computer (in press), IGARSS 2025 · Code: SEE-Unet, Tiny-UCloudnet

All-in-One Image Restoration

Remote sensing imagery used in environmental monitoring is often affected by blur, noise and atmospheric haze, which can reduce the reliability of downstream analysis. Most all-in-one restoration methods rely on explicit degradation guidance and are evaluated mainly on natural or aerial images. We proposed VCP-IR, a visual-conditioned framework that extracts a global visual descriptor from the degraded input image to generate adaptive prompt features within a Transformer-based encoder–decoder. This enables restoration of multiple degradation types without explicit degradation labels or task-specific branches. VCP-IR is evaluated on both aerial and ground-based sky imagery, which are widely used in Earth observation, cloud monitoring and solar irradiance forecasting.

Paper: Signal Processing: Image Communication (in press) · Code: VCP-IR

Architecture of VCP-IR: a Transformer-based encoder-decoder that uses a visual descriptor of the degraded image to generate prompts, turning a degraded aerial image into a restored one.

Solar Forecasting

Solar power is highly intermittent, so accurate short-term estimates and forecasts of solar irradiance are essential for integrating photovoltaic (PV) generation into energy grids. We develop deep learning models that estimate and forecast irradiance from infrared sky images, including MobileNetV2-based CNN regression, hybrid CNN–LSTM models and multiscale deep feature fusion, and we study how input window size and forecasting horizon affect multi-horizon forecasting.

Because these models are meant to run on low-cost, low-power hardware next to the camera, we also explore ways to make them smaller and faster. One approach is knowledge distillation, where a compact “student” network learns to mimic a larger, more accurate “teacher” network, keeping much of its accuracy at a fraction of the size. We have also investigated model compression through pruning.

Ongoing: we are studying how temporal design choices and model scaling affect short-term solar forecasting.

Papers: see the Publications page · Code: solar-irradiance-estimation-dl, Solar_Estimation_KD, SOLAR_FORECASTING_KD, MDFF, solar-forecasting-pruning, SolarForecasting-DesignTradeoffs

Knowledge distillation for solar irradiance estimation: a sky image feeds a large teacher model and a compact student model, and the student is trained on both the true target and the teacher's softened predictions.

Other Projects

Secure ECG Monitoring with Federated Learning and Blockchain. Continuous ECG monitoring is crucial for early heart disease prediction, but traditional data management and model training can expose sensitive patient information. We built a secure ECG monitoring system that integrates federated learning with blockchain technology. A mobile app helps patients monitor their heart condition and receive feedback from doctors, all connected through a private blockchain server, while federated learning trains the ECG classification model across patient devices. In simulation, the federated learning system scaled to 500 clients.

Paper: CAPSI 2025, link · Code: Federated-ECG

System architecture of the secure ECG monitoring system: patient and doctor apps, an ECG processing API, a private blockchain storing doctor feedback, and federated learning that updates local and global models.

Page updated

Google Sites

Report abuse