Early detection of wildfires is crucial to mitigate their catastrophic effects on lives and natural and built environments. Smoke (referring to fire smoke in this thesis) detection is considered a promising approach for early fire (EF) detection since smoke plumes are usually the first indicators when wildfires occur. Using satellite imagery for smoke detection stands out because it provides cost-effective monitoring that covers large scales and remote areas. Scene-level detection classifies imagery into smoke or other classes based on whether smoke is in the imagery. It offers faster inference and better scalability than pixel-level detection, which aims to identify all smoke pixels individually. This makes scene-level detection ideal for fire disaster mitigation. In this thesis, smoke detection refers to satellite-based scene-level smoke detection unless specified.

Despite significant advancements in smoke detection enabled by deep learning (DL), several limitations in previous research still need to be addressed. Existing DL models are often complex and excessively demanding in terms of power supply, memory usage, and computing resources. Such models are difficult to deploy onboard satellites, particularly small satellites (SmallSats). Additionally, previous DL-based smoke detection research focused on RGB imagery and did not pay enough attention to spectral information that could potentially improve detection accuracy. Furthermore, an effective mechanism for fast model development for multiple satellites is yet to be investigated to enhance the timely detection of EF smoke. This needs to consider that new sensors usually lack observational data and that imagery data from different sensors often present significant disparities.

This thesis develops three innovative approaches to address these limitations progressively. To facilitate the study, two multispectral imagery training datasets, Landsat6c and Sentinel7c with a medium spatial resolution of 30 metres and 10 metres respectively, are created. Landsat6c additionally incorporates one near-infrared (NIR) and two shortwave infrared (SWIR) bands and Sentinel7c includes four additional infrared (IR) bands (two NIR and two SWIR ), compared to the sole publicly available training dataset USTC SmokeRS, derived from Moderate Resolution Imaging Spectroradiometer (MODIS) three-band (RGB) imagery with a low spatial resolution (1 km).

The first approach proposes a lightweight convolutional neural network (CNN) model called Variant Input Bands for Smoke Detection (VIB SD). VIB SD contains less than 2% of the parameters of the state-of-the-art model SAFA (1.66 million versus 84.2 million) but demonstrates competitive accuracy (93.57% versus 96.22%) when trained using USTC SmokeRS. This approach proceeds to train VIB SD using Landsat6c with five different band combinations to investigate the contributions of IR bands to detection accuracy. Results show that incorporating the NIR band enhances accuracy compared to using solely RGB bands (84.82% versus 83.20%) and integrating both SWIR bands leads to further improvements compared to using just one SWIR band (86.45% versus 85.64%). Case studies illustrate VIB SD’s effectiveness in detecting EF smoke amidst cloud cover when trained with Landsat6c.

The second approach further investigates how to effectively explore useful spectral information in IR bands to improve smoke detection accuracy. Specifically, this approach introduces a DL module named Input Amplification (IA) which enables DL models to automatically learn class-oriented spectral patterns. IA amplifies the input band dimension (e.g., three or six) to 32 (determined through experiments), with the learned spectral patterns added as pseudo bands. This allows for simultaneous learning of multiple spectral patterns and integrating them with the original bands. Functioning as an input pre-processing block, IA facilitates seamless integration with various DL architectures. The effectiveness of IA is demonstrated through its integration with different CNN architectures (i.e., ResNet50, InceptionResNetV2, MobileNetV2, and VIB SD) and testing on both USTC SmokeRS and Landsat6c. Significant accuracy improvements were observed for the examined CNN models after integration with IA, showcasing IA’s great potential in advancing smoke detection. Notably, higher accuracy increments were obtained when using Landsat6c featuring additional IR bands (4.61%, 1.08%, 1.9%, and 3.54% respectively for ResNet50, InceptionResNetV2, MobileNetV2, and VIB SD).

The third approach introduces a novel cross-sensor transfer learning method, notably aided by the IA module. This aims to facilitate fast model development for multiple satellites, considering data disparities across different sensors and the limited observational data available from new sensors. Landsat6c, originally containing 1836 images (about 600 images per class), was expanded to 2770 images (more than 900 images per class) and used as the source domain. Sentinel7c, consisting of only 351 images (about 120 images per class), served as the target domain. The model incorporating IA and VIB SD, named IA VIB SD, was employed. The transferability of the Landsat6c-trained IA VIB SD model to Sentinel7c was investigated using various transfer learning techniques and compared to the performance of the benchmark IA VIB SD model exclusively trained on pure Sentinel7c data. The proposed transfer learning method resulted in a transferred model with an average accuracy 5% higher than the benchmark model. Notably, the proposed transfer learning method outperformed conventional transfer learning methods by more than 1% in terms of accuracy, even when trained on only 10% of the Sentinel7c dataset.

In summary, the core achievement of this thesis is the IA VIB SD model. This model significantly enhances smoke detection accuracy by employing both RGB and IR bands and learning class-oriented spectral patterns from these bands. It has great potential to facilitate onboard satellite smoke detection, particularly on SmallSats and SmallSat constellations, due to its lightweight design. Aided by the IA module’s capability of adapting learned spectral patterns, IA VIB SD pretrained on one sensor demonstrates high accuracy when transferred to a new sensor using minimal training data from the target sensor.