Very High-Resolution Forest Mapping with TanDEM-X InSAR Data and Self-Supervised Learning

TL;DR

This work tackles the challenge of mapping forests at a very high 6 m resolution using TanDEM-X InSAR data with limited labeled references. It proposes a self-supervised learning framework based on a convolutional autoencoder, trained with an inpainting pretext task, to initialize a U-Net for downstream forest mapping, and compares it to a fully-supervised baseline. Across Pennsylvania and the Amazon, the SSL-In E+D approach achieves competitive or superior performance with far fewer labeled samples, and significantly improves edge delineation of forests and narrow features such as roads and clear-cuts. The findings demonstrate the practicality of SSL for spaceborne InSAR-based forest mapping and offer a path toward large-scale, high-resolution forest products in data-sparse regions.

Abstract

Deep learning models have shown encouraging capabilities for mapping accurately forests at medium resolution with TanDEM-X interferometric SAR data. Such models, as most of current state-of-the-art deep learning techniques in remote sensing, are trained in a fully-supervised way, which requires a large amount of labeled data for training and validation. In this work, our aim is to exploit the high-resolution capabilities of the TanDEM-X mission to map forests at 6 m. The goal is to overcome the intrinsic limitations posed by midresolution products, which affect, e.g., the detection of narrow roads within vegetated areas and the precise delineation of forested regions contours. To cope with the lack of extended reliable reference datasets at such a high resolution, we investigate self-supervised learning techniques for extracting highly informative representations from the input features, followed by a supervised training step with a significantly smaller number of reliable labels. A 1 m resolution forest/non-forest reference map over Pennsylvania, USA, allows for comparing different training approaches for the development of an effective forest mapping framework with limited labeled samples. We select the best-performing approach over this test region and apply it in a real-case forest mapping scenario over the Amazon rainforest, where only very few labeled data at high resolution are available. In this challenging scenario, the proposed self-supervised framework significantly enhances the classification accuracy with respect to fully-supervised methods, trained using the same amount of labeled data, representing an extremely promising starting point for large-scale, very high-resolution forest mapping with TanDEM-X data.

Figures (9)

• Figure 1: TanDEM-X acquisitions used in this study over the the Pennsylvania state, USA. (a) Ground coverage and (b) Total $h_{\mathrm{amb}}$ distribution. The individual colors show the relative patch contribution of the different years to the total.
• Figure 2: CNN architectures used in the study: (a) Convolutional autoencoder (CAE) and (b) U-Net. The structure of the encoding part is common to both CNNs.
• Figure 3: Considered TanDEM-X acquisitions when using (a) 22% and (b) 1.5% of the available labeled data over Pennsylvania. They are used for training the downstream task of forest mapping.
• Figure 4: $F^w_1$-score for the investigated approaches over Pennsylvania testing area for the test subsets: (a) Short $h_\mathrm{amb}$; (b) Mid $h_\mathrm{amb}$; (c) Large $h_\mathrm{amb}$; The results are presented for the different amount of considered labeled data in the supervised learning part: 1.5%, 8% and 22% (horizontal axis). The baseline approach consists of a fully-supervised training using 100% of the available TanDEM-X images with labeled data. FSL corresponds to the same architecture of the baseline (U-Net) trained in a fully-supervised manner with less labeled data. The other cases correspond to SSL pretext task (SSL-Id: identity reconstruction with CAE, SSL-In: inpainting with masked CAE), followed by a downstream task of forest mapping (E+D: encoder and decoder initialized from the SSL pretext task and then trained, D: encoder weights frozen from SSL pretext task and decoder trained).
• Figure 5: Map view of the confusion matrices values for 4 different areas on patches of 1024 $\times$ 1024 pixels. On the ground truth plots on the left-hand side, green areas correspond to forests and white areas to non-forested zones. Beside the baseline case, the models are trained with 1.5% of the labeled data. The rows correspond to areas inside images acquired in ascending orbit direction in 2011 and 2012 for the different test subsets: (a) Short $h_{\textrm{amb}}$, (b) Mid $h_{\textrm{amb}}$, and (c) Large $h_{\textrm{amb}}$. The fourth row (d) is part of a TanDEM-X image acquired in 2013 in descending orbit direction and with a $h_{\textrm{amb}} = 85~m$.