Comparison of Data Fusion Methods Using Crowdsourced Data in Creating a Hybrid Forest Cover Map

Myroslava Lesiv,Elena Moltchanova,Steffen Fritz,Linda See,Alexis Comber,Dmitry Schepaschenko,Anatoly Shvidenko

doi:10.3390/rs8030261

Abstract

Data fusion represents a powerful way of integrating individual sources of information to produce a better output than could be achieved by any of the individual sources on their own. This paper focuses on the data fusion of different land cover products derived from remote sensing. In the past, many different methods have been applied, without regard to their relative merit. In this study, we compared some of the most commonly-used methods to develop a hybrid forest cover map by combining available land cover/forest products and crowdsourced data on forest cover obtained through the Geo-Wiki project. The methods include: nearest neighbour, naive Bayes, logistic regression and geographically-weighted logistic regression (GWR), as well as classification and regression trees (CART). We ran the comparison experiments using two data types: presence/absence of forest in a grid cell; percentage of forest cover in a grid cell. In general, there was little difference between the methods. However, GWR was found to perform better than the other tested methods in areas with high disagreement between the inputs.

Highlights

Land cover maps provide useful information on the geographical distribution of different land cover types, as well as on land cover change over time
We extend the work of Schepaschenko et al [7], who used only geographically-weighted logistic regression (GWR) to create their hybrid forest maps, by considering other commonly-used methods of data fusion for creating a global forest map
logistic regression (LR) is statistically-significantly lower than NB, but otherwise, the four methods (NN, classification and regression trees (CART), NB and LR) appear to perform fairly

Summary

Introduction

Land cover maps provide useful information on the geographical distribution of different land cover types, as well as on land cover change over time. Land cover products are widely used as input data in various applications, such as climate change models, management of natural resources, environmental monitoring and comprehensive spatial quantification of ecosystems and landscapes, among many others [1]. Maps of forest cover, in particular, provide valuable inputs to a diverse range of applications, including the modelling of forest growth and productivity, the assessment of bioenergy potentials, carbon flux monitoring and REDD+. The overall trend has been towards higher spatial resolution, such as the 30-m maps of the percentage of forest cover, forest cover gain and loss by Hansen [2]. The 30-m Globeland product [3,4] These maps were developed from Landsat high resolution satellite imagery, which has only been made possible because this data stream has recently become freely available [5]. Hansen’s forest cover map could be considered an exception, as it is Remote Sens. 2016, 8, 261; doi:10.3390/rs8030261 www.mdpi.com/journal/remotesensing

Methods

Discussion

Conclusion