While Landsat has proved to be effective for monitoring many elements of forest condition and change, the platform has well-documented limitations in measuring forest structure, the vertical distribution of the canopy. This is important because structure determines several key ecosystem functions, including: carbon storage; habitat suitability; and timber volume. Canopy structure is directly measured by LiDAR, and it should be possible to train Landsat structure models at a highly local scale with the dense, global sample of full waveform LiDAR observations collected by NASA’s Global Ecosystem Dynamics Investigation (GEDI). Local models are expected to perform better because: (a) such models may take advantage of localized correlations between structure and canopy surface reflectance; and (b) to the extent that models revert to the mean of the calibration data due to a lack of discrimination, local models will revert to a more representative mean. We tested Landsat-based relative height predictions using a new GEDI asset on Google Earth Engine, described here. Mean prediction error declined by 23% and important prediction biases at the extremes of the range of canopy height dropped as model calibration became more local, minimizing forest structure signal saturation commonly associated with Landsat and other passive optical sensors. Our results suggest that Landsat-based maps of structural variables such as height and biomass may substantially benefit from the kind of local calibration that GEDI’s dense sample of LiDAR data supports.

Population growth rates in Sub-Saharan East Africa are among the highest in the world, creating increasing pressure for land cover conversion. To date, however, there has been no comprehensive assessment of regional land cover change, and most long-term trends have not yet been quantified. Using a designed sample of satellite-based observations of historical land cover change, we estimate the areas and trends in nine land cover classes from 1998 to 2017 in Ethiopia, Kenya, Uganda, Malawi, Rwanda, Tanzania, and Zambia. Our analysis found an 18,154,000 (±1,580,000) ha, or 34.8%, increase in the area of cropland in East Africa. Conversion occurred primarily from Open Grasslands, Wooded Grasslands, and Open Forests, causing a large-scale reduction in woody vegetation classes. We observed far more conversion (by approximately 20 million hectares) of woody classes to less-woody classes than succession in the direction of increasing trees and shrubs. Spatial patterns within our sample highlight regional land cover conversion hotspots, such as the Central Zambezian Miombo Woodlands, as potential areas of concern related to the conservation of natural ecosystems. Our findings reflect a rapidly growing population that is moving into new areas, with a 43.5% increase in the area of Settlements over the three-decade period. Our results show the areas and ecoregions most impacted by three decades of human development, both spatially and statistically.

Disturbance monitoring is an important application of the Landsat times series, both to monitor forest dynamics and to support wise forest management at a variety of spatial and temporal scales. In the last decade, there has been an acceleration in the development of approaches designed to put the Landsat archive to use towards these causes. Forest disturbance mapping has moved from using individual change-detection algorithms, which implement a single set of decision rules that may not apply well to a range of scenarios, to compiling ensembles of such algorithms. One approach that has greatly reduced disturbance detection error has been to combine individual algorithm outputs in Random Forest (RF) ensembles trained with disturbance reference data, a process called stacking (or secondary classification). Previous research has demonstrated more robust and sensitive detection of disturbance using stacking with both multialgorithm ensembles and multispectral ensembles (which make use of a single algorithm applied to multiple spectral bands). In this paper, we examined several additional dimensions of this problem, including: (1) type of algorithm (represented by processes using one image per year vs. all historical images); (2) spectral band choice (including both the basic Landsat reflectance bands and several popular indices based on those bands); (3) number of algorithm/spectral-band combinations needed; and (4) the value of including both algorithm and spectral band diversity in the ensembles. We found that ensemble performance substantially improved per number of model inputs if those inputs were drawn from a diversity of both algorithms and spectral bands. The best models included inputs from both algorithms, using different variants of shortwave-infrared (SWIR) and near-infrared (NIR) reflectance. Further disturbance detection improvement may depend upon the development of algorithms which either interrogate SWIR and NIR in new ways or better highlight disturbance signals in the visible wavelengths.

Background: Crown fires are often intense and fast spreading and hence can have serious impacts on soil, vegetation, and wildlife habitats. Fire managers try to prevent the initiation and spread of crown fires in forested landscapes through fuel management. The minimum fuel conditions necessary to initiate and propagate crown fires are known to be strongly influenced by four stand structural variables: surface fuel load (SFL), fuel strata gap (FSG), canopy base height (CBH), and canopy bulk density (CBD). However, there is often a lack of quantitative data about these variables, especially at the landscape scale. Methods: In this study, data from 123 sample plots established in pure, even-aged, Pinus radiata and Pinus pinaster stands in northwest Spain were analyzed. In each plot, an intensive field inventory was used to characterize surface and canopy fuels load and structure, and to estimate SFL, FSG, CBH, and CBD. Equations relating these variables to Sentinel-2A (S-2A) bands and vegetation indices were obtained using two non-parametric techniques: Random Forest (RF) and Multivariate Adaptive Regression Splines (MARS). Results: According to the goodness-of-fit statistics, RF models provided the most accurate estimates, explaining more than 12%, 37%, 47%, and 31% of the observed variability in SFL, FSG, CBH, and CBD, respectively. To evaluate the performance of the four equations considered, the observed and estimated values of the four fuel variables were used separately to predict the potential type of wildfire (surface fire, passive crown fire, or active crown fire) for each plot, considering three different burning conditions (low, moderate, and extreme). The results of the confusion matrix indicated that 79.8% of the surface fires and 93.1% of the active crown fires were correctly classified; meanwhile, the highest rate of misclassification was observed for passive crown fire, with 75.6% of the samples correctly classified. Conclusions: The results highlight that the combination of medium resolution imagery and machine learning techniques may add valuable information about surface and canopy fuel variables at large scales, whereby crown fire potential and the potential type of wildfire can be classified.