Land cover classification of mountainous environments continues to be a challenging remote sensing problem,owing to landscape complexities exhibited by the region.This study explored a multiple classifier system(MCS)a...Land cover classification of mountainous environments continues to be a challenging remote sensing problem,owing to landscape complexities exhibited by the region.This study explored a multiple classifier system(MCS)approach to the classification of mountain land cover for the Khumbu region in the Himalayas using Sentinel-2 images and a cloud-based model framework.The relationship between classification accuracy and MCS diversity was investigated,and the effects of different diversification and combination methods on MCS classification performance were comparatively assessed for this environment.We present ten MCS models that implement a homogeneous ensemble approach,using the high performing Random Forest(RF)algorithm as the selected classifier.The base classifiers of each MCS model were developed using different combinations of three diversity techniques:(1)distinct training sets,(2)Mean Decrease Accuracy feature selection,and(3)‘One-vs-All’problem reduction.The base classifier predictions of each RFMCS model were combined using:(1)majority vote,(2)weighted argmax,and(3)a meta RF classifier.All MCS models reported higher classification accuracies than the benchmark classifier(overall accuracy with 95% confidence interval:87.33%±0.97%),with the highest performing model reporting an overall accuracy(±95% confidence interval)of 90.95%±0.84%.Our key findings include:(1)MCS is effective in mountainous environments prone to noise from landscape complexities,(2)problem reduction is indicated as a stronger method over feature selection in improving the diversity of the MCS,(3)although the MCS diversity and accuracy have a positive correlation,our results suggest this is a weak relationship for mountainous classifications,and(4)the selected diversity methods improve the discriminability of MCS against vegetation and forest classes in mountainous land cover classifications and exhibit a cumulative effect on MCS diversity for this context.展开更多
The use of freely-available multi-source imagery for mapping vegetation in montane terrain is important for many developing countries that do not have the funding for high-resolution data capture.Radar images are also...The use of freely-available multi-source imagery for mapping vegetation in montane terrain is important for many developing countries that do not have the funding for high-resolution data capture.Radar images are also now freely available and include Sentinel-1 in dual polarisation,and PALSAR-2.These images can penetrate cloud cover and provide the advantage of acquiring data in a cloudy tropical region.This research evaluated whether the addition of radar with optical and topographic data improves classification accuracy in a montane region in Sri Lanka.Six classification experiments were designed based on different combinations of image data to test whether radar data improved land cover classification accuracy compared with optical data alone.Random forest classifier in the Google Earth Engine has been utilised to classify the tropical montane vegetation.The results indicate that radar or optical data alone cannot obtain satisfactory results.However,when combining radar with optical data the overall accuracy increased by approximately 5%,and by an additional 2%when topography data were added.The highest accuracy(92%)was achieved with multiple imagery,and adding the vegetation indices improved the model slightly by 0.3%.In addition,feature importance analysis showed that radar data makes a significant contribution to the classification.These positive outcomes demonstrate that freely-accessible multi-source remotely-sensed data have impressive capability for vegetation mapping,and support the monitoring and managing of forest ecological resources in tropical montane regions.展开更多
文摘Land cover classification of mountainous environments continues to be a challenging remote sensing problem,owing to landscape complexities exhibited by the region.This study explored a multiple classifier system(MCS)approach to the classification of mountain land cover for the Khumbu region in the Himalayas using Sentinel-2 images and a cloud-based model framework.The relationship between classification accuracy and MCS diversity was investigated,and the effects of different diversification and combination methods on MCS classification performance were comparatively assessed for this environment.We present ten MCS models that implement a homogeneous ensemble approach,using the high performing Random Forest(RF)algorithm as the selected classifier.The base classifiers of each MCS model were developed using different combinations of three diversity techniques:(1)distinct training sets,(2)Mean Decrease Accuracy feature selection,and(3)‘One-vs-All’problem reduction.The base classifier predictions of each RFMCS model were combined using:(1)majority vote,(2)weighted argmax,and(3)a meta RF classifier.All MCS models reported higher classification accuracies than the benchmark classifier(overall accuracy with 95% confidence interval:87.33%±0.97%),with the highest performing model reporting an overall accuracy(±95% confidence interval)of 90.95%±0.84%.Our key findings include:(1)MCS is effective in mountainous environments prone to noise from landscape complexities,(2)problem reduction is indicated as a stronger method over feature selection in improving the diversity of the MCS,(3)although the MCS diversity and accuracy have a positive correlation,our results suggest this is a weak relationship for mountainous classifications,and(4)the selected diversity methods improve the discriminability of MCS against vegetation and forest classes in mountainous land cover classifications and exhibit a cumulative effect on MCS diversity for this context.
文摘The use of freely-available multi-source imagery for mapping vegetation in montane terrain is important for many developing countries that do not have the funding for high-resolution data capture.Radar images are also now freely available and include Sentinel-1 in dual polarisation,and PALSAR-2.These images can penetrate cloud cover and provide the advantage of acquiring data in a cloudy tropical region.This research evaluated whether the addition of radar with optical and topographic data improves classification accuracy in a montane region in Sri Lanka.Six classification experiments were designed based on different combinations of image data to test whether radar data improved land cover classification accuracy compared with optical data alone.Random forest classifier in the Google Earth Engine has been utilised to classify the tropical montane vegetation.The results indicate that radar or optical data alone cannot obtain satisfactory results.However,when combining radar with optical data the overall accuracy increased by approximately 5%,and by an additional 2%when topography data were added.The highest accuracy(92%)was achieved with multiple imagery,and adding the vegetation indices improved the model slightly by 0.3%.In addition,feature importance analysis showed that radar data makes a significant contribution to the classification.These positive outcomes demonstrate that freely-accessible multi-source remotely-sensed data have impressive capability for vegetation mapping,and support the monitoring and managing of forest ecological resources in tropical montane regions.
