Remote Sensing | Interview with One of the Issue Cover Authors—Dr. Jose Bermudez

Dr. Jose Bermudez is one of the authors of the Cover Paper featured in Volume 18, Issue 15, entitled “An Integrated Deep Learning Approach to Estimate Canopy Height and Uncertainty by Combining Seasonal Optical, SAR and Limited GEDI LiDAR Data over Ontario’s Managed Forests, Canada”, published in Remote Sensing (ISSN 2072-4292).

Based on the positive evaluations from the reviewers and Academic Editors of Dr. Jose Bermudez‘s group’s article, we have selected their article as the Issue Cover for display on the Remote Sensing website.

The following is an interview with Dr. Jose Bermudez:

1. Congratulations on your published paper! Could you please briefly introduce yourself, the main research content and the key findings of your study?
I am Jose Bermudez, the corresponding author of this paper and a researcher at McMaster University’s School of Earth, Environment & Society. My work focuses on remote sensing and deep learning for monitoring forest structure.
In this study, my co-authors and I developed a probabilistic deep-learning framework to map forest canopy height across Ontario at 30 m resolution. The approach combines seasonal Landsat optical imagery, Sentinel-1 C-band SAR, annual ALOS-PALSAR-2 L-band SAR, and sparse canopy-height measurements from NASA’s GEDI spaceborne LiDAR. It produces both a continuous canopy-height map and a pixel-level uncertainty map.
Our primary model achieved an R² of 0.70, an RMSE of 3.65 m, and a bias of only 0.23 m when evaluated against independent airborne LiDAR. Replacing the commonly used Gaussian loss with a more outlier-robust Laplace loss reduced RMSE by 13.7% and decreased systematic bias from 1.43 m to 0.14 m under matched conditions. Seasonal observations also improved accuracy substantially, while even the SAR-only models outperformed four evaluated global canopy-height products across every forest class in our Ontario test area.

2. Forest ecosystems play a critical role in regulating the global carbon cycle. Why is accurate mapping of forest canopy height so important for understanding forests and their carbon dynamics?
Canopy height is one of the most informative remotely sensed indicators of forest structure. It is closely related to tree size and aboveground biomass, so accurate height maps help us estimate how much carbon forests store. Canopy height is also relevant to carbon sequestration because larger trees generally continue to accumulate substantial amounts of carbon.
Accurate mapping is particularly important because even moderate systematic errors in height can propagate into large errors in biomass and carbon-stock estimates. Spatially continuous height maps can also reveal differences among forest types, identify tall or regenerating stands, support forest inventories, and help monitor disturbances and recovery. However, a canopy-height map alone is not sufficient: users also need to know where its predictions are reliable. That is why our study provides a calibrated uncertainty layer alongside the height estimates.

3. Large-scale forest monitoring faces challenges because satellite LiDAR provides valuable 3D information but only samples limited locations. How does your approach overcome this limitation?
GEDI provides highly valuable three-dimensional measurements, but it collects discrete footprints rather than continuous wall-to-wall observations. In our study, we use those GEDI footprints as training labels for a ResUNet-based deep-learning model. The model learns relationships between canopy height and spatially continuous satellite observations from Landsat, Sentinel-1, and ALOS-PALSAR-2. Once trained, it transfers the structural information contained in the GEDI samples to all pixels covered by the other satellite datasets, including areas between GEDI tracks.
The network also uses spatial context over a relatively large receptive field, rather than treating each pixel as an isolated observation. This allows it to learn from patterns such as forest texture, patch structure, edges, and neighbourhood heterogeneity. Finally, we use an ensemble of models and a probabilistic Laplace formulation to estimate uncertainty at each pixel. This is important because extrapolating from sparse LiDAR samples inevitably introduces uncertainty, and the resulting confidence map helps users identify predictions that require caution.

4. Your study integrates optical remote sensing, synthetic aperture radar (SAR) and spaceborne LiDAR data. What does each type of observation contribute, and why is combining different satellite data important?
Each observation provides a different view of the forest:

• Landsat optical data captures spectral properties of vegetation, including greenness, canopy condition, and seasonal phenology. Winter, summer, and fall composites help distinguish leaf-on, leaf-off, and snow or freeze-related conditions.
• Sentinel-1 C-band SAR is sensitive mainly to leaves, smaller branches, canopy surfaces, moisture, and freeze-thaw conditions. It operates through clouds and therefore provides reliable observations in cloud-prone regions.
• ALOS-PALSAR-2 L-band SAR has a longer wavelength and penetrates more deeply into the canopy, making it more responsive to larger branches and trunks.
• GEDI LiDAR supplies direct vertical measurements of canopy structure. Although spatially sparse, these measurements provide the height labels needed to train the model.

Combining these datasets is important because their strengths are complementary. Optical data may be unavailable under persistent cloud and can saturate in dense forests. Radar works independently of daylight and most cloud conditions, while different radar wavelengths interact with different parts of the canopy. Our results show that multi-sensor fusion was the strongest driver of overall accuracy. Seasonal Landsat and Sentinel-1 carried most of the predictive information, with ALOS-PALSAR-2 providing an additional improvement.

5. Looking ahead, what advances in satellite technology, artificial intelligence, or data availability do you think will have the greatest impact on global forest monitoring?
I expect several developments to have a major impact. First, new SAR missions will provide more frequent observations at longer wavelengths. Data from ALOS-4 and NISAR, together with the P-band BIOMASS mission, should improve our ability to characterize dense forests where optical and shorter-wavelength radar signals tend to saturate.
Second, combining multiple spaceborne LiDAR missions will be very valuable. For example, integrating GEDI with ICESat-2 could increase the density and geographic coverage of structural training data, particularly at high latitudes beyond GEDI’s observational limit.
Third, advances in artificial intelligence should help models learn across regions, sensors, and time while making better use of spatial and seasonal context. Self-supervised Earth-observation foundation models, transfer learning, and uncertainty-aware deep learning are especially promising. Explainable-AI methods will also be important because operational users need to understand why a model produces a particular estimate and when that estimate should be trusted.
Finally, the greatest benefits will come when satellite imagery, reference measurements, computing resources, trained models, and validation datasets are openly available. That will make it possible to update forest maps more frequently and evaluate whether methods transfer reliably across different biomes.

6. How do you think the open access method of publishing impacts authors?
Open access publishing significantly increases the visibility and practical impact of research. Because readers do not face a subscription barrier, the paper can reach scientists, students, forest managers, government agencies, Indigenous and local organizations, and researchers in countries or institutions with limited journal access.
For authors, this can increase readership, collaboration, reuse, and potentially citations. It also supports transparency and reproducibility. In our case, open publication complements the public availability of the satellite data, processing pipeline, model-training scripts, trained model weights, and final canopy-height and uncertainty maps. This allows other researchers to evaluate the method, adapt it to other regions, and build upon our results.