Deep convolutional regression modelling for forest parameter retrieval

Abstract

<p>Accurate forest monitoring is crucial as forests are major global carbon sinks. Additionally, accurate prediction of forest parameters, such as forest biomass and stem volume (SV), has economic importance. Therefore, the development of regression models for forest parameter retrieval is essential. <p>Existing forest parameter estimation methods use regression models that establish pixel-wise relationships between ground reference data and corresponding pixels in remote sensing (RS) images. However, these models often overlook spatial contextual relationships among neighbouring pixels, limiting the potential for improved forest monitoring. The emergence of deep convolutional neural networks (CNNs) provides opportunities for enhanced forest parameter retrieval through their convolutional filters that allow for contextual modelling. However, utilising deep CNNs for regression presents its challenges. One significant challenge is that the training of CNNs typically requires continuous data layers for both predictor and response variables. While RS data is continuous, the ground reference data is sparse and scattered across large areas due to the challenges and costs associated with in situ data collection. <p>This thesis tackles challenges related to using CNNs for regression by introducing novel deep learning-based solutions across diverse forest types and parameters. To address the sparsity of available reference data, RS-derived prediction maps can be used as auxiliary data to train the CNN-based regression models. This is addressed through two different approaches. <p>Although these prediction maps offer greater spatial coverage than the original ground reference data, they do not ensure spatially continuous prediction target data. This work proposes a novel methodology that enables CNN-based regression models to handle this diversity. Efficient CNN architectures for the regression task are developed by investigating relevant learning objectives, including a new frequency-aware one. To enable large-scale and cost-effective regression modelling of forests, this thesis suggests utilising C-band synthetic aperture radar SAR data as regressor input. Results demonstrate the substantial potential of C-band SAR-based convolutional regression models for forest parameter retrieval.