Author: Oleg Antropov
The ‘Improved Semisupervised UNet Deep Learning Model for Forest Height Mapping With Satellite SAR and Optical Data’ paper was recently published in IEEE Journal of Selected Topics in Applied Earth Observation and Remote Sensing. Although the work was conducted mostly outside the context of the Forest Carbon Monitoring (FCM) project, the analysis provides understanding on potential future development directions of the carbon monitoring platform.
The paper, coordinated by a FCM team member Oleg Antropov, compares Multiple Linear Regression (MLR) and Random Forest (RF) approaches tested in the FCM, to a deep learning approaches. Same type of Sentinel-1 and Sentinel-2 data is used. In the paper, an improved semi-supervised deep learning model is introduced, and its suitability for modeling the relationship between forest structural variables and satellite Copernicus remote sensing imagery is demonstrated.
The developed CPRSeUNet model is based on a popular UNet model, modified and fine-tuned to improve the forest parameter prediction performance. Within the improved model, squeeze-and-excitation blocks are embedded to recalibrate the multisource features via retrieved channel-wise self-attention and a novel cross-pseudo regression strategy is implemented to train the model in a semi-supervised way. The improvement imposes consistency learning on two perturbed network branches: 1) generating regression pseudo-reference; 2) expanding the dataset size.
For demonstration of the model and assessment of its performance versus traditional prediction approaches, satellite synthetic aperture radar (SAR) Sentinel-1 and multi-spectral optical Sentinel-2 images are used as remote sensing data, and forest tree height calculated from airborne laser scanner (ALS) data as a reference. The study area of size 50 km x 50 km is located in a boreal forestland in Central Finland.
The proposed approach shows higher accuracy compared to traditional machine learning methods such as random forests and boosting trees, and baseline UNet model. Best accuracy figures for forest tree height are achieved with combined SAR and optical imagery and are as small as 24.1% root-mean-square error (RMSE) on pixel-level and 15.4% RMSE on forest stand level.
To the best of our knowledge, this is the first application of deep-learning models in the context of forest inventory mapping using a combination of SAR and optical satellite data. The model is apparently capable to extract features contributing to removal of “signal saturation” observed for high forest with larger biomass, as shown in Figure 1.
Approaches like the one introduced in this paper offer a potential direction for further improvement of the system to be set-up and demonstrated in the Forest Carbon Monitoring project. Particularly the greater resistance against saturation is a valuable characteristic in areas with high volume forests. Widening the selection of methods available in the platform in the future will improve our capability to meet the varying user needs in varying conditions.
Figure 1 . Forest height prediction performance for various EO datasets and prediction methods. (Ge et al. 2022)
Full Citation: S. Ge, H. Gu, W. Su, J. Praks and O. Antropov, “Improved Semisupervised UNet Deep Learning Model for Forest Height Mapping With Satellite SAR and Optical Data,” in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 15, pp. 5776-5787, 2022, doi: 10.1109/JSTARS.2022.3188201