Forest Height Inversion by Convolutional Neural Networks Based on L-Band PolInSAR Data Without Prior Knowledge Dependency

Dandan Li, Hailiang Lu, Chao Li, Linda Mohaisen, Weipeng JingIEEE JSTARS 2023 · OpenReview

Abstract

Forest height is a key forest parameter which is of great significance for monitoring forest resources, calculating forest biomass, and observing the global carbon cycle. Because the PolInSAR system could provide various object information including height, shape and direction sensitivity, and spatial distribution, it becomes a powerful means for measuring forest height. The proposed framework utilizes deep learning and builds upon traditional DEM differencing and coherence amplitude inversion algorithms. By using L band PolInSAR data, a new convolutional neural network (CNN) model is established in which the estimated results of DEM differencing and coherence amplitude inversion are used as labels. Furthermore, the PCGrad optimization strategy is used for updating the gradient automatically in the training stage. This model could not only build a relationship between complex coherence and forest height but also makes full use of the spatial context information by using the CNN layers. Experiments are carried out based on the simulated data and real data, named Lope forest site, which are collected by uninhabited aerial vehicle synthetic aperture radar in the NASA AfriSAR campaign. Compared to the classic forest height inversion algorithms, the proposed framework has achieved a higher level of accuracy and performance on RMSE (10.15 m) and $R^{2}$ (0.87). Overall, the proposed framework does not require LiDAR data as prior knowledge and can be performed on various forest scenes. Consequently, it will hopefully serve as a useful approach for improvements in forest height inversion based on PolInSAR data.

Introduction

Measuring forest height is critical for calculating biomass and monitoring the global carbon cycle. While LiDAR (Light Detection and Ranging) provides highly accurate height maps, its deployment is often limited by high costs and an inability to penetrate cloud cover. Polarimetric Interferometric Synthetic Aperture Radar (PolInSAR) offers a robust alternative, capable of characterizing vertical forest structures regardless of weather conditions. Specifically, L-band radar is preferred for forest height estimation due to its superior ability to penetrate the canopy and interact with the ground layer, providing a more complete vertical profile compared to shorter wavelengths like X or C bands.

Comparison of radar band penetration in forest canopies

Traditional PolInSAR inversion models, such as the Random Volume over Ground (RVoG) model, often require complex parameterization or external prior knowledge. While deep learning has emerged as a promising solution, most existing models rely on expensive LiDAR data for training labels. This research addresses this dependency by proposing a framework that eliminates the need for LiDAR-derived training data.

Methodology

The researchers developed the Polarimetric Interferometric Deep Learning-based Framework (PIDLF). The core innovation is a training strategy that utilizes "noisy" labels from two traditional algorithms with known, opposing biases:

  1. DEM Differencing: Known to consistently underestimate forest height.
  2. Coherence Amplitude Inversion (Sinc): Known to consistently overestimate forest height.
The proposed Polarimetric Interferometric Deep Learning-based Framework (PIDLF)

The framework employs a three-layer CNN that processes complex coherences from HH, HV, and VV channels. To maintain spatial resolution, pooling layers were omitted. To handle the conflicting biases of the training labels, the team implemented the Projected Conflicting Gradients (PCGrad) optimization strategy. This multi-task learning technique identifies and resolves gradient conflicts, allowing the model to converge on a height estimate that effectively "averages out" the systematic errors of the traditional methods.

Results and Discussion

The PIDLF was validated using simulated data and real-world data from the NASA AfriSAR campaign at the Lope forest site in Gabon. Using LVIS LiDAR (RH100) as a validation standard, the framework achieved a Root Mean Square Error (RMSE) of 10.15m and an $R^2$ of 0.87.

This performance significantly outperformed traditional DEM differencing (RMSE 17.19m) and the complex RVoG model (RMSE 10.35m). Qualitatively, the CNN-generated height maps successfully captured the complex spatial textures of the tropical rainforest, demonstrating the benefits of incorporating spatial context through convolutional layers. The model effectively learned the underlying relationship between radar coherence and height, outperforming the very algorithms used to train it.

Conclusion

By removing the dependency on LiDAR for training, the PIDLF offers a portable and scalable solution for forest height inversion. While an RMSE of 10m in dense tropical forests leaves room for further refinement, the ability of a neural network to outperform its biased "teachers" marks a significant advancement in remote sensing. Future work will explore more efficient architectures and the integration of additional coherence data to further enhance precision across diverse forest biomes.

Summary updated

Abstract

Forest height is a key forest parameter which is of great significance for monitoring forest resources, calculating forest biomass, and observing the global carbon cycle. Because the PolInSAR system could provide various object information including height, shape and direction sensitivity, and spatial distribution, it becomes a powerful means for measuring forest height. The proposed framework utilizes deep learning and builds upon traditional DEM differencing and coherence amplitude inversion algorithms. By using L band PolInSAR data, a new convolutional neural network (CNN) model is established in which the estimated results of DEM differencing and coherence amplitude inversion are used as labels. Furthermore, the PCGrad optimization strategy is used for updating the gradient automatically in the training stage. This model could not only build a relationship between complex coherence and forest height but also makes full use of the spatial context information by using the CNN layers. Experiments are carried out based on the simulated data and real data, named Lope forest site, which are collected by uninhabited aerial vehicle synthetic aperture radar in the NASA AfriSAR campaign. Compared to the classic forest height inversion algorithms, the proposed framework has achieved a higher level of accuracy and performance on RMSE (10.15 m) and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$R^{2}$</tex-math></inline-formula> (0.87). Overall, the proposed framework does not require LiDAR data as prior knowledge and can be performed on various forest scenes. Consequently, it will hopefully serve as a useful approach for improvements in forest height inversion based on PolInSAR data.
BibTeX
@article{li2023forest,
  author = {Dandan Li and Hailiang Lu and Chao Li and Linda Mohaisen and Weipeng Jing},
  title = {{Forest Height Inversion by Convolutional Neural Networks Based on L-Band PolInSAR Data Without Prior Knowledge Dependency}},
  journal = {IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens.},
  year = {2023},
  url = {https://openreview.net/forum?id=AQOkfaP0P4},
}
Figure