Fiona Cawkwell

Temporal optimisation of image acquisition for land cover classification with random forest and MODIS time-series

The analysis and classification of land cover is one of the principal applications in terrestrial remote sensing. Due to the seasonal variability of different vegetation types and land surface characteristics, the ability to discriminate land cover types changes over time. Multi-temporal classification can help to improve the classification accuracies, but different constraints, such as financial restrictions or atmospheric conditions, may impede their application. The optimisation of image acquisition timing and frequencies can help to increase the effectiveness of the classification process. For this purpose, the Feature Importance (FI) measure of the state-of-the art machine learning method Random Forest was used to determine the optimal image acquisition periods for a general (Grassland, Forest, Water, Settlement, Peatland) and Grassland specific (Improved Grassland, Semi-Improved Grassland) land cover classification in central Ireland based on a 9-year time-series of MODIS Terra 16 day composite data (MOD13Q1). Feature Importances for each acquisition period of the Enhanced Vegetation Index (EVI) and Normalised Difference Vegetation Index (NDVI) were calculated for both classification scenarios. In the general land cover classification, the months December and January showed the highest, and July and August the lowest separability for both VIs over the entire nine-year period. This temporal separability was reflected in the classification accuracies, where the optimal choice of image dates outperformed the worst image date by 13% using NDVI and 5% using EVI on a mono-temporal analysis. With the addition of the next best image periods to the data input the classification accuracies converged quickly to their limit at around 8–10 images. The binary classification schemes, using two classes only, showed a stronger seasonal dependency with a higher intra-annual, but lower inter-annual variation. Nonetheless anomalous weather conditions, such as the cold winter of 2009/2010 can alter the temporal separability pattern significantly. Due to the extensive use of the NDVI for land cover discrimination, the findings of this study should be transferrable to data from other optical sensors with a higher spatial resolution. However, the high impact of outliers from the general climatic pattern highlights the limitation of spatial transferability to locations with different climatic and land cover conditions. The use of high-temporal, moderate resolution data such as MODIS in conjunction with machine-learning techniques proved to be a good base for the prediction of image acquisition timing for optimal land cover classification results.

On The Glaciers of Bylot Island, Nunavut, Arctic Canada

ABSTRACT The present extent of glacier ice on Bylot Island, Arctic Canada, is mapped using high-resolution Landsat 7 ETM+ satellite imagery. The island is 43% ice covered, with 4783 km2 of ice. Most ice is centered on the northwest-southeast–trending Byam Martin Mountains, flowing outward as radial valley glaciers and piedmont lobes. The largest glacier is 49 km long and 6.5 km wide. The majority of glaciers terminate on land, but many have margins ending in lakes and two calve into the sea. The late summer snowline, mapped from satellite imagery, is highest along the southern and central parts of the island at about 1050 m, with lower values along the east-northeastern margin of the ice down to about 700 m. These snowline-elevation differences suggest a predominant moisture source from the northeast. Several valley glaciers and piedmont lobes have deformed medial moraines and ice-surface foliation suggesting past surge activity. Ten glaciers are interpreted to be of possible surge-type. The modern extent of glaciers is compared with that of two earlier time intervals. First, we have mapped glacier margins in several areas of Bylot Island from aerial photographs acquired in 1958 and 1961. Secondly, former positions of ice fronts are mapped from moraine systems deposited during the Neoglacial maximum and identified on satellite data. Glaciers have retreated from 0.9 to 1.8 km since the Neoglacial maximum about 120 years ago, with most retreat occurring between 1958/1961 and 2001. Approximately 253 km2 or 5% of the 1958/1961 ice-covered area has been lost. Overall, marked glacier retreat has occurred, although a few glaciers, possibly of surge-type, show small readvances. This retreat is consistent with observed climate warming in the Canadian Arctic, especially since the 1960s.

Spatial and temporal variability in the snowpack of a High Arctic ice cap: implications for mass-change measurements

Abstract Interpretation of ice mass elevation changes observed by satellite altimetry demands quantification of the proportion of elevation change which is attributable to variations in firn densification. Detailed stratigraphic logging of snowpack structure and density was carried out at ~1km intervals along a 47 km transect on Devon Ice Cap, Canada, in spring (pre-melt) and autumn (during/ after melt) 2004 and 2006 to characterize seasonal snowpack variability across the full range of snow facies. Simultaneous meteorological measurements were gathered. Spring (pre-melt) snowpacks show low variability over large spatial scales, with low-magnitude changes in density. The end-of-summer/ autumn density profiles show high variability in both 2004 and 2006, with vastly different melt regimes generating dissimilar patterns of ice-layer formation over the two melt seasons. Dye-tracing experiments from spring to autumn 2006 reveal that vertical and horizontal distribution of meltwater flow within and below the annual snowpack is strongly affected by the pre-existing, often subtle stratigraphic interfaces in the snowpack, rather than its bulk properties. Strong interannual variability suggests that using a simple relationship between air temperature, elevation and snowpack densification to derive mass change from measurements of elevation change across High Arctic ice caps may be misguided. Melt timing and duration are important extrinsic factors governing snowpack densification and ice-layer formation in summer, rather than averaged air temperatures.

Remote sensing of recent glacier changes in the Canadian Arctic

The Canadian Arctic contains the largest area of land ice (~150,000 km2) on Earth outside the ice sheets of Greenland and Antarctica and is a potentially significant contributor to global sea level change. The current ice cover includes large ice caps that are remnants of the Wisconsinan Laurentide and Innuitian ice sheets, and many smaller ice caps and valley glaciers that formed during the late Holocene. Most of these ice masses have decreased in area over the past century as a result of climate warming in the first half of the 20th century and since the mid-1980s. In general, smaller ice masses have lost a higher proportion of their area, but the largest total area losses have come from the larger ice caps. Both iceberg calving and negative surface mass balances have contributed to this episode of glacier shrinkage. Long-term calving rates are not well known, however, and many tidewater glaciers exhibit velocity variability on a range of timescales that may affect calving rates. Floating ice shelves in northern Ellesmere Island have lost over 90 % of their area in the 20th century, with the most recent phase of disintegration occurring since 2000. Some fjords in the region are now ice free for the first time in over 3000 years. Regional rates of mass loss have accelerated strongly since 2005, and Canadian Arctic glaciers and ice caps have emerged as the most significant non–ice sheet contributor to the nonsteric component of global sea level rise.

Mass balance of the Prince of Wales Icefield, Ellesmere Island, Nunavut, Canada

[1]xa0This paper estimates the mass balance of the Prince of Wales Icefield, Ellesmere Island, Canada, averaged over four decades, from measurements of surface mass balance (SMB) and iceberg calving. Shallow ice core net accumulation measurements and annual mass balance stake measurements are used in conjunction with a digital elevation model and knowledge of the location of the dominant moisture source for precipitation over the ice cap to interpolate and extrapolate spatial patterns of SMB across the Prince of Wales Icefield. The contribution of iceberg calving to the mass balance is calculated from estimates of (1) the annual volume of ice discharged at the major tidewater glacier termini and (2) the annual volume loss or gain due to terminus fluctuations. Two different approaches to determining the SMB conclude that the SMB of the ice field is approximately in balance (average equals −0.1 ± 0.4 km3 w.e. a−1, where w.e. means water equivalent) largely because of its proximity to the main year-round moisture source that is the Smith Sound portion of the North Open Water polynya. Iceberg calving is a highly significant component of mass loss (−1.9 ± 0.2 km3 w.e. a−1) and is sufficient to make the overall mass balance of the ice field averaged over the period 1963–2003 clearly negative (−2 ± 0.45 km3 w.e. a−1, equivalent to a mean-specific mass balance across the ice field of −0.1 m w.e. a−1). The Prince of Wales Icefield contributes ∼0.005 mm a−1 to global eustatic sea level rise.

Determination of cloud top amount and altitude at high latitudes

Cloud identification over snow and ice has proved to be a difficult process to automate due to the similarity in their visible and thermal properties. A method is described here which utilises the nadir and forward views of the Along Track Scanning Radiometer to determine the height of the surface observed to a nominal accuracy of ±1000 m (pixel resolution). A digital elevation model allows the surface topography to be ascertained, with the remainder of high elevation features classed as cloud. Visual verification of the resulting cloud masks indicates that the stereo-matcher identifies 10–20% more cloud than is recognised by the human eye, but nearly 100 % of cloud-free land identified by stereo-matching was confirmed by examination of the images. Further validation of the cloud top heights using radiosonde data, indicates 73% of the stereo-matched heights to be within 500 m of the radiosonde predicted cloud tops.

Modeling Managed Grassland Biomass Estimation by Using Multitemporal Remote Sensing Data—A Machine Learning Approach

More than 80% of agricultural land in Ireland is grassland, which is a major feed source for the pasture based dairy farming and livestock industry. Many studies have been undertaken globally to estimate grassland biomass by using satellite remote sensing data, but rarely in systems like Irelands intensively managed, but small-scale pastures, where grass is grazed as well as harvested for winter fodder. Multiple linear regression (MLR), artificial neural network (ANN) and adaptive neuro-fuzzy inference system (ANFIS) models were developed to estimate the grassland biomass (kg dry matter/ha/day) of two intensively managed grassland farms in Ireland. For the first test site (Moorepark) 12 years (2001–2012) and for second test site (Grange) 6 years (2001–2005, 2007) of in situ measurements (weekly measured biomass) were used for model development. Five vegetation indices plus two raw spectral bands (RED=red band, NIR=Near Infrared band) derived from an 8-day MODIS product (MOD09Q1) were used as an input for all three models. Model evaluation shows that the ANFIS (

Cattle stocking rates estimated in temperate intensive grasslands with a spring growth model derived from MODIS NDVI time-series

R_{{rm{Moorepark}}}^2 = ;0.85,;;{rm{RMS}}{{rm{E}}_{{rm{Moorepark}}}} = ;11.07

Application of statistical and machine learning models for grassland yield estimation based on a hypertemporal satellite remote sensing time series

;

Upland vegetation mapping using Random Forests with optical and radar satellite data

R_{{rm{Grange}}}^2 = ;0.76,;;{rm{RMS}}{{rm{E}}_{{rm{Grange}}}} = ;15.35