Machine Learning Reveals Two Decades of Shrinking Greenery Around Malawi’s Largest Lagoon

Deep in Nkhotakota District, along the western shoreline of Lake Malawi, lies Chia Lagoon, the country’s largest lagoon and a newly designated Key Biodiversity Area. A team of Malawian researchers has now assembled the most detailed satellite-based picture yet of how the landscape surrounding this ecologically vital wetland has transformed over twenty years, and the results reveal a catchment under sustained pressure. Using machine-learning classification of Landsat imagery from 2004, 2014, and 2024, the study found that mapped vegetation in the roughly 3,367-square-kilometre catchment fell from 519 square kilometres, about 15.4 percent of the area, to just 330 square kilometres, or 9.8 percent, by 2024. That net loss of 189 square kilometres represents more than a third of the green cover that existed when the record began.

The research, published in BMC Environmental Science, applied a Random Forest supervised classification algorithm, an ensemble machine-learning method that builds hundreds of decision trees and merges their predictions through majority voting. Random Forest has become a workhorse of satellite-based land-cover mapping because it handles complex, non-linear relationships among spectral variables and remains robust when training data are noisy. In this study, the team ran 500 trees per classification, with the number of predictor variables evaluated at each split set by the square-root rule recommended in the foundational literature. Training samples, roughly 150 per land-cover class for each observation year, were drawn from GPS-based field surveys and visual interpretation of high-resolution imagery, then divided 70:30 between training and independent validation.

The classification scheme was deliberately simple: three classes covering water, vegetation, and bare land. That choice reflects a practical constraint of working with 30-metre-resolution Landsat pixels in a heterogeneous landscape, where finer categories become spectrally indistinguishable. Water captured the lagoon itself, its channels, and permanently inundated areas. Vegetation included riparian strips, woodland patches, shrubs, and wetland plants. Bare land was the broadest bucket of all, encompassing exposed soil, harvested and fallow fields, degraded surfaces, sparsely vegetated ground, and even small built-up features, which could not be separated from soil at this resolution. The authors are explicit that changes in this class therefore signal shifts in non-vegetated surface rather than any single process such as settlement growth or deforestation.

To keep the sensors comparable across two decades, the team used Landsat 5 Thematic Mapper imagery for 2004 and Landsat 8 Operational Land Imager data for 2014 and 2024, relying on Collection 2 Level-2 surface reflectance products and only the spectrally equivalent reflective bands: blue, green, red, near-infrared, and two shortwave-infrared channels. All three scenes were acquired in August, during the dry season, to minimise the confounding effects of seasonal vegetation phenology and wetland inundation. Cloud and shadow masking, geometric alignment, and resampling to a common 30-metre grid completed the preprocessing, which was carried out in ENVI 5.3 with spatial analysis performed in ArcGIS Pro 3.4.

The accuracy results strengthened steadily across the record. Overall classification accuracy rose from 82.0 percent in 2004 to 90.0 percent in 2014 and 96.0 percent in 2024, with a pooled accuracy of 89.3 percent across 450 independent validation samples. Water was classified with consistently high reliability, with producer’s accuracy between 96 and 98 percent, while confusion concentrated where it usually does in medium-resolution work: between vegetation and bare land, particularly in transitional zones of sparse or regrowing cover. Rather than reporting raw pixel counts, the researchers adopted the probabilistic area-estimation framework of Olofsson and colleagues, adjusting mapped areas with 95 percent confidence intervals and propagating uncertainty through the change calculations. This is an important methodological discipline, because unadjusted pixel tallies can overstate the precision of landscape-change estimates.

The decadal breakdown tells a story in two contrasting phases. Between 2004 and 2014, mapped vegetation collapsed from 519 square kilometres to 261 square kilometres, a loss of roughly 258 square kilometres and 7.6 percentage points of catchment share, while bare land surged from 80.2 percent to 87.7 percent of the landscape. Between 2014 and 2024, however, the trend partially reversed: vegetation recovered to 330 square kilometres, a gain of about 69 square kilometres, and bare land eased back to 86.4 percent. Even so, the 2024 vegetation extent remained well below the 2004 baseline, meaning the partial recovery only clawed back around a quarter of the earlier loss. Mapped water extent, meanwhile, declined across the full period, from 147.6 square kilometres in 2004 to 128.35 square kilometres in 2024, a net reduction of about 19 square kilometres.

Two spectral indices provided complementary context for these patterns. The Normalized Difference Vegetation Index, which exploits the contrast between red-band absorption and near-infrared reflectance to gauge vegetation greenness, showed mean values rising slightly from 0.21 in 2004 to 0.25 in 2024, with the spread of values narrowing over time. The Modified Normalized Difference Water Index, which distinguishes water from other surfaces using the contrast between green reflectance and shortwave-infrared absorption, showed mean values drifting from negative 0.05 to negative 0.02 across the same span. The authors are careful to note that because these indices were derived from the same Landsat scenes used for classification, they serve as supporting spectral evidence rather than independent validation of the mapped area changes, and that pixel-level index values do not translate directly into class-area estimates.

What the study deliberately does not do is assign causes. The bare-land category bundles together too many distinct surface types, and three dry-season snapshots a decade apart cannot capture seasonal variability, short-term hydrological fluctuation, or the interplay of agricultural expansion, climate variability, and land management that likely drives change in the catchment. The authors state plainly that their results should be read as mapped land-cover changes rather than direct evidence of ecological degradation or recovery. Establishing drivers would require temporally matched field, socioeconomic, climatic, and land-management data. Residual sensor differences between Landsat 5 and Landsat 8, along with uncertainty in retrospective interpretation of historical reference imagery, further bound the interpretation, even though surface reflectance standardisation and comparable band selection were used to mitigate them.

Even with those caveats, the findings carry real weight for conservation. Chia Lagoon functions as a semi-closed extension of Lake Malawi, supporting threatened endemic fish species, wetland birds such as the black heron and lesser masked weaver, and the livelihoods of roughly 7,857 households dependent on fisheries and smallholder agriculture. Its designation as a Key Biodiversity Area in 2024 underscores its significance, and the surrounding vegetation plays a crucial role in regulating erosion, sediment delivery, and nutrient movement into the wetland. The study’s uncertainty-adjusted maps offer exactly the kind of spatially explicit baseline needed to prioritise restoration sites, evaluate management interventions, and report against international commitments under the Convention on Biological Diversity and Sustainable Development Goal 15. The work was funded through the UK government’s Darwin Initiative, and the authors call for future research that pairs machine-learning-based monitoring with field-based ecological, hydrological, and socioeconomic measurements, an integration that would transform these satellite snapshots into a genuine understanding of why Malawi’s largest lagoon catchment is changing and what can be done to safeguard it.

Subject of Research: Machine-learning-based remote sensing analysis of land-use and land-cover change in the Chia Lagoon catchment, Malawi, from 2004 to 2024

Article Title: Machine learning analysis of land use change dynamics in Malawi’s Chia Lagoon

Article References: Machine learning analysis of land use change dynamics in Malawi’s Chia Lagoon. (n.d.). https://doi.org/10.1186/s44329-026-00065-7

Image Credits: AI Generated

DOI: 10.1186/s44329-026-00065-7

Keywords: Chia Lagoon, Malawi, land-use change, Random Forest, Landsat, remote sensing, wetland conservation, NDVI, MNDWI, Key Biodiversity Area, machine learning, freshwater ecosystems