Exploring the potential of freely available remote sensing data and machine learning to monitor forest above ground biomass in the Miombo woodlands of Zambia

Kanja, Kennedy and Atkinson, Peter (2025) Exploring the potential of freely available remote sensing data and machine learning to monitor forest above ground biomass in the Miombo woodlands of Zambia. PhD thesis, Lancaster University.

Text (2025KKanjaPhD)
PhD_Thesis_-_KKanja_35487875_.pdf - Published Version
Restricted to Repository staff only until 17 November 2026.
Download (5MB)

Abstract

Tropical dry forests play a critical role in carbon storage, biodiversity conservation, and supporting livelihoods. Yet, they remain among the most poorly studied and monitored ecosystems globally. In Zambia, the Miombo woodlands, a dominant tropical dry forest type, provide a wide range of ecosystem services that support local communities and are central to national climate goals, particularly under the frameworks of REDD+ and forest landscape restoration (FLR). Accurate, spatially explicit, and temporally consistent monitoring of forest above-ground biomass (AGB) is vital for carbon accounting, policy implementation, and understanding landscape change. However, mapping and monitoring AGB in Miombo woodlands is challenging due to limited field data, driven by the high cost and logistical complexity of field-based data collection. The heterogeneous nature of these woodlands, resulting from both natural and anthropogenic disturbances, as well as ongoing forest regrowth, further complicates their mapping using remote sensing techniques. This thesis explores the potential of freely available remote sensing data and machine learning methods to monitor forest AGB in the Miombo woodlands of Zambia. Specifically, it integrates open-access Synthetic Aperture Radar (SAR) and optical satellite data with advanced computational approaches, including U-Net convolutional neural networks (CNNs) and Random Forest (RF) models, to improve forest type classification and AGB estimation. A hierarchical classification approach is proposed, using RF for land use/land cover (LULC) mapping and U-Net for discriminating Miombo woodlands into functional forest types. The RF model achieved an overall LULC classification accuracy of 93%, with forest class F1-scores ranging from 93% to 96% across different seasons. The U-Net CNN effectively delineated Miombo woodlands into reference, degraded, and regrowth forest types, achieving F1-scores of 85%, 73%, and 72%, respectively. An extensive National Forest Inventory (NFI) dataset was used to assess the capability of multi-seasonal SAR (Sentinel-1) and optical (Landsat-8) imagery in mapping AGB. The best performing AGB model combined SAR backscatter from the hot season, optical bands from the dry season, and vegetation indices from the hot season, yielding R = 0.69, MAE = 14.01 Mg ha⁻¹, and RMSE = 18.23 Mg ha⁻¹. Model performance varied across Zambia’s agro-ecological zones (R = 0.44–0.79). A time-series analysis of AGB from 2014 to 2021 revealed spatially variable trends in biomass gain and loss, underscoring the potential of combining single-date inventory and free satellite data for near-real-time forest monitoring. This research contributes to the development of cost-effective and scalable methodologies for national forest monitoring systems. The findings confirm that freely available satellite sensor data and limited inventory data can be used to generate accurate, repeatable biomass monitoring to support REDD+ and other conservation efforts.