detection of harvested forest areas in italy using landsat imagery
TRANSCRIPT
lable at ScienceDirect
Applied Geography 48 (2014) 102e111
Contents lists avai
Applied Geography
journal homepage: www.elsevier .com/locate/apgeog
Detection of harvested forest areas in Italy using Landsat imagery
P. Borrelli a,b,*, S. Modugno c, P. Panagos a, M. Marchetti d, B. Schütt b, L. Montanarella a
a Joint Research Centre of the European Commission, Institute for Environment and Sustainability, Via E. Fermi, 2749, I-21027 Ispra, VA, ItalybDepartment of Earth Sciences, Physical Geography, Freie Universität Berlin, Malteserstrabe 74-100, Haus H, 12249 Berlin, GermanycCentre for Landscape and Climate Research, University of Leicester, Bennett Building, University Road, Leicester LE1 7RH, UKdUniversity of Molise, Department of BioScience and Territory, Isernia, Italy
Keywords:Forest monitoringClear-cut mappingForest inventoryPolicy making
* Corresponding author. Joint Research Centre ofInstitute for Environment and Sustainability, Via E. FeItaly. Tel.: þ39 0332 789072; fax: þ39 0332 786394.
E-mail addresses: [email protected](P. Borrelli).
0143-6228/$ e see front matter � 2014 Published byhttp://dx.doi.org/10.1016/j.apgeog.2014.01.005
a b s t r a c t
This study presents a thorough approach, based on the application of multi-spectral remote sensingLandsat imagery, to determine human-induced forest cover change in Italy during the decade2002e2011. A total of 785.6 �1 04 ha of forestland was mapped using the main forest classes describedwithin the CORINE land cover 2006 database (3.11 e broad-leaved forest; 3.12 e coniferous forest; 3.13 e
mixed forest). The approach employs multi-temporal Landsat imagery to determine large-scale spatio-temporal variations in forest cover with a high degree of precision. The semi-automated procedure isbased on Normalized Difference Vegetation Index (NDVI) pixel-oriented image differencing technique.The results were validated and rectified as a result of on-screen visual interpretation, whereby all thefalse-positive forest changes that were incorrectly mapped during the automatic procedure were iden-tified and removed. The derived high-resolution data of forest cover change show that 317,535 ha (4.04%of the total forest area in Italy) were harvested during the period under review. The 125,272 individualclear-cut areas identified are mainly located within protected areas of the European Natura 2000network. The outcome of this study is a publicly accessible database that can encourage further studies inthe framework of international biodiversity and soil protection conventions (http://eusoils.jrc.ec.europa.eu/library/themes/erosion/italy/). The methodology can contribute to the monitoring of human-inducedforest changes in support of the Kyoto Protocol.
� 2014 Published by Elsevier Ltd.
1. Introduction
The side effects of deforestation and tree harvesting are majorenvironmental issues on a global scale (FAO, 2012). The world’sforestland is cleared, degraded and fragmented by timber har-vesting, man-made fires and land-use conversion (Cochrane, 2003;Richards & Tucker, 1988; Williams, 2000). Every year, about 13million hectares (ha) of forest are converted to other land uses (FAO,2010). Agriculture is still a primary cause of deforestation, ac-counting for 80% of deforestation, followed by logging (14%) andfuel wood (5%) (UNFCCC, 2007).
In Europe, the largest clearance of forest area occurred betweenthe Classical period and the Industrial Revolution (Kaplan, 2009).
the European Commission,rmi, 2749, I-21027 Ispra, VA,
pa.eu, [email protected]
Elsevier Ltd.
Currently, logging is still widely practised in most of the EUMember States (European Commission, 2003). It is estimated thatabout 420 million m3 of roundwood forests were harvested in theEuropean Union (EU-27) in 2010 to meet the domestic timber de-mand (Eurostat, 2011).
Despite the amount of Eurostat data available about the pro-duction and trading of wood in the European Union, no informationis available on the locations of tree extractions. This knowledgedeficit could trigger serious issues, particularly given the fact that avast part of the European forest area is privately owned (e.g. 72% ofthe 368,820 km2 of forestland of Central-Western Europe is pri-vately owned) and widely exploited for wood supply. Moreover, theintense harvesting of forests goes against the forest managementprogrammes of the European Commission (e.g. forest resourcestocks, productivity and harvesting activities e MCPFE, 2009).
Remote sensing observation data can help address this knowl-edge deficit. Remote sensing (RS) has been used as a powerful toolfor detecting changes in land-use and vegetation over the pastdecades (Coppin & Bauer, 1996; Kennedy & Spies, 2004; Mas, 1999).
Fig. 1. Study site (in green the target forest areas according to CORINE land cover 2006e EEA, 2011). [For interpretation of colour referred in this figure legend, the reader isreferred to web version of the article.]
Table 1Italian forestland statistics by NUTS-2 administrative units (INFC, 2007).
NUTS-2 Forest Other woodedlands [ha]
Total forestcoverage
Piedmont 870,594 69,522 940,116Valle d’Aosta 98,439 7489 105,928Lombardy 606,045 59,657 665,703South Tyrol 336,689 35,485 372,174Trentino 375,402 32,129 407,531Veneto 397,889 48,967 446,856Friuli V.G. 323,832 33,392 357,224Liguria 339,107 36,027 375,134Emilia Romagna 563,263 45,555 608,818Tuscany 1,015,728 135,811 1,151,539Umbria 371,574 18,681 390,255Marche 291,394 16,682 308,076Lazio 543,884 61,974 605,859Abruzzi 391,492 47,099 438,590Molise 132,562 16,079 148,641Campania 384,395 60,879 445,274Apulia 145,889 33,151 179,040Basilicata 263,098 93,329 356,426Calabria 468,151 144,781 612,931Sicily 256,303 81,868 338,171Sardinia 583,472 629,778 1,213,250National coverage 8,759,200 1,708,333 10,467,533
P. Borrelli et al. / Applied Geography 48 (2014) 102e111 103
RS is a recognised technique for monitoring forest change at mul-tiple scales (Collins &Woodcock, 1996; Hansel et al., 2008; Potapov,Turubanova, & Hansen, 2011; Singh, 1989), as RS data includecomprehensive information about the spatial and temporal distri-bution of forest changes (Coppin & Bauer, 1996). The mostfrequently used data for assessing forest cover change are thosecaptured by Landsat satellite sensors (MSS, TM, ETMþ), which areavailable since 1972 (Asner, Keller, Pereira, & Zweede, 2002; Coppin& Bauer, 1994; Cohen et al., 2002; Coops et al., 2010; Hansen et al.,2001; Healey, Cohen, Zhiqiang, & Krankina, 2005; Li et al., 2009;Wilson & Sader, 2002; Woodcock, Macomber, Pax-Lenney, &Cohen, 2001; among others). The widespread use of the Landsatsensor data is attributable to the good compromise that they offerin terms of spatial resolution and temporal coverage (Williams,2006). In addition, a broad database of orthorectified Landsatscenes is available free of charge from the United States GeologicalSurvey (USGS) (Woodcock et al., 2008). Landsat imagery isconsidered to be the most cost-efficient tool for forest monitoring,with respect to high-resolution images. The cost of obtaining high-resolution images for a large study area can be prohibitive, rangingfrom V1 to V2/km (SPOT5) to V15 to V20/km (Ikonos, QuickBird).However, given the small average size of the Italian clear-cut areas(about 2.5 ha e Borrelli, Rondón, & Schütt, 2013a), the coarse-resolution imagery of other sensors (e.g. MODIS e 250 � 250 m,NOAA-AVHRR e 1.1�1.1 km) is of limited use for the current study.
Post-classification analysis is one of the most commonly appliedmethodologies for detecting environmental changes (Lu, Mausel,Brondízio, & Moran, 2004; Rodriguez-Galiano & Chica-Olmo,2012). However, the potential errors deriving from the classifica-tion of land units (Linke et al., 2009) may represent an unnecessaryrestriction during the detection of forest cover change. Therefore,the cover of large forestlands could be better monitored bymethods that rely on pixel-oriented change detection techniques(Singh, 1989) and use vegetation indices (Lyon, Yuan, Lunetta, &Elvidge, 1998; Wilson & Sader, 2002). Borrelli et al. (2013a) havetested a pixel-oriented change detection technique over a34,000 km2 area of central Italy, with encouraging results (KappaIndex of Agreement: 0.906).
The objective of this study is to compile an inventory of forestcover changes at national-scale based on an image differencingtechnique (Singh, 1989). The study area is the territory of Italy,which has a total forestland area covering 104.7 � 105 ha (34.7% ofthe national surface) (INFC, 2007). The resulting database coversthe decade 2002e2011. It can further promote studies in thecontext of international conventions on areas such as biodiversityprotection (European Commission, 2011), soil conservation(European Commission, 2006) and compliance with the re-quirements of the Kyoto Protocol.
2. Study area
The study area (Fig. 1) covers about 785.6 � 104 ha, corre-sponding to the main forest units of the CORINE land cover 2006database (EEA, 2011), i.e. broad-leaved forests (547.9� 104 ha, 70%),coniferous forests (128.6 � 104 ha, 17%) and mixed forests(109.1 � 104 ha, 13%).
The dominant tree species are Quercus (petraea, robur, petraea,cerris, carpinifolia, sativa, ilex) Fagus sylvatica, Picea abies and Abiesalba (Vacchiano, Magnani, & Collalti, 2012). On average, 34.7% of theItalian territory is covered by forests. The coefficient of woodiness(INFC, 2005) is lower in the southern regions (the EU NUTS-2administrative units of Apulia, Basilicata, Calabria) and on theislands (Sicily and Sardinia) (Table 1), where other forms of woodedland (e.g. shrubs and macchia) represent a substantial portion ofthe forest area. The most densely wooded regions are Liguria and
Trentino, with a coverage rate of 69.7% and 65.5% respectively. Animportant aspect of the Italian forest area is its ownership. A total of63.5% of the forest area (forest and other wooded land) is privatelyowned, 32.4% is publicly owned, while almost 4% of the area has notbeen classified (Pompei & Gasparini, 2007). About 366.3 � 103 ha(41.8% of the Italian forestland) is currently managed as coppiceforest (INFC, 2007).
P. Borrelli et al. / Applied Geography 48 (2014) 102e111104
The elevation of forestland in Italy ranges from 0 m a.s.l. at thelowest points (coastline) up to 2675 m a.s.l. in the Lombardy Alps.The mean elevation is 800 m a.s.l. The forestland surface is domi-nated by the Apennine Mountains and largely features moun-tainous (60%) and hilly (26%) sections, with only 14% of the forestarea being found on flat plains. The forestland is characterised bymoderate slopes with an average gradient of 21� (51.6% of Italianforestland is found on slopes with a gradient greater than 20�). Thegeology is significantly heterogeneous, characterised by complexlithological types (APAT, 2005). A great variety of soil types isidentified (Rusco, Filippi, Marchetti, & Montanarella, 2003)although medium-texture Cambisols dominate the Italian land-scape (Panagos, Van Liedekerke, Jones, & Montanarella, 2012). Thedominant climates are warm Mediterranean, temperate and sub-continental (Cs and Cf of the Köppen-Geiger climate classificationsystem). Annual precipitation varies from 350 mm yr�1 in thesouthern coastal areas of Sardinia and Sicily to 2500e3500 mm yr�1 in the Carnic Alps region. The national average isabout 970 mm yr�1. Average annual temperatures range from 5 to10 �C (typical of the Apennines and Alps highlands) to 14e16 �C(along the southern coasts).
3. Methodology
3.1. Image acquisition and pre-processing
Using GloVis (http://glovis.usgs.gov), the following Landsatimagery was selected and downloaded from the Earth ResourcesObservation and Science Center (EROS) of the United StatesGeological Survey (USGS): 30-m Landsat Thematic Mapper (TM)imagery (175 scenes) and 30-m Landsat Enhanced Thematic Map-per Plus (ETMþ) imagery (231 scenes). This imagery allows for thebest coverage of the study area over a time period of 10 years. Onlyimages taken during summer months were selected, as (i) theimages acquired during the dry season are clear of atmospherichaze and so have fewer inter-image differences between the sun’sangle and azimuth, atmospheric transmissions and soil moisture(Williams, 2006); (ii) harvesting operations are generally onlypermitted from November to March/April in Italy. Most of theLandsat data used are standard level-one terrain-corrected (L1T)(USGS, 2011). These kinds of images have not been atmosphericallycorrected, but are geometrically adjusted to remove systematicgeometric errors related to the positions of the sensor, satellite andEarth (USGS, 2011). Accordingly, all the scenes were radiometricallycorrected to remove atmospheric attenuation in order to addressatmospheric scattering effects. The ENVI 4.7 software was used forthis purpose. Here, the digital number (DN) values of each imagewere converted to surface reflectance (Williams, 2006) and weresubject to normalisation by dark object subtraction (DOS) as rec-ommended by Chavez and Mackinnon (1994). This kind of imagenormalisation improves the accuracy of the classification andcalculation of forest cover loss (Hansen et al., 2008; Potapov,Hansen, Stehman, Loveland, & Pittman, 2008).
3.2. Masking operations
After the pre-processing procedure, a forest/non-forest maskwas generated to identify the changes in forestland. CORINEshapefiles (EEA, 2011) of the ‘broad-leaved forest’ (code 3.11),‘coniferous forest’ (code 3.12) and ‘mixed forest’ (code 3.13) landuse units were selected, merged, converted into raster format(30 � 30 m) and reprojected using the Landsat data coordinatesystem (i.e. UTM WGS84 Fuse 33). The Landsat images then un-derwent two masking operations: (i) the forested areas wereextracted during the cover change detection phase (ENVI 4.7) using
the forest/non-forest mask; (ii) clouds, shadows, or sporadic noisedetected in the imagery were masked out.
3.3. Forest cover change detection
The acquired Landsat images were qualitatively classified anddivided into two databases depending on their missing data (cloudand shadow cover, noise and data gaps generated by the Landsat 7Scan Line Corrector-Off (SLC-Off) malfunction). The main database(database A) contains those images with minimum levels ofdisturbance, while the remaining images were filed in ‘database B’.In the following, ‘database B’ serves as a back-up system to fill anyinformation gaps in database A.
Once the images were pre-processed and classified, an imagedifferencing technique (Singh, 1989) was applied to assess thedetection of forest cover change (Fig. 2). This method subtracts thespatially registered satellite data of two images using a pixel-by-pixel procedure. Satellite vegetation indices are generallyemployed for this purpose, as they can characterise the type,amount and condition of vegetation data acquired by the sensors(Huete et al., 2002). Forest cover change studies based on vegeta-tion indices have been applied worldwide with satisfactory results(Singh, 1989; Lyon et al., 1998; Elmore, Mustard, Manning, & Lobell,2000). In addition to single-band information (i.e. red-band, NI-Band, MI-band), the most commonly used vegetation indices arethe Normalized Difference Vegetation Index e NDVI (Rouse, 1973),the Normalized Difference Water Index e NDWI (McFeeters, 1996),the Normalized Difference Moisture Index e NDMI (Zha, Gao, & Ni,2003), the Enhanced Vegetation Index e EVI (Liu & Huete, 1995),and the Leaf Area Index e LAI (Chen & Cihlar, 1996). Using the re-sults obtained in a study of areas in central Italy (Borrelli et al.,2013a), where many of the aforementioned indices wereemployed, we make use of the Normalized Difference VegetationIndex (NDVI ¼ (near infrared band e red band)/(near infraredband þ red band); Jensen, 1986). The NDVI index proved to be veryaccurate for the detection of forest cover change in Italy (97.6% ofthe algorithm-derived clear-cut areas match the interpreted clear-cut areas >0.45 ha).
Changes in forest cover were calculated by subtracting the NDVIvalues of a Landsat image from those of the preceding year. Negativevalues indicated areas that had been subjected to tree-harvestingactivities or fires. In some cases, it was necessary to carry outrepeated change detection procedures using Landsat TM images ofother years (‘not preceding year’) in order to fill the gaps in LandsatETMþ SLC-Off data and other images that had cloud, shadow ornoise interference. These additional operations were necessarybecause some information gaps could not be filled using the imagesof ‘database B’ or overlapping data of neighbouring Landsat images.As a result, forest changes were partially documented (only forzones with no overlaps) on a two-yearly basis for the followingyears: Path/Row 193/28, years 2005e2006 and 2008e2009; Path/Row 193/29, year 2008e2009; Path/Row 188/33, year 2005e2006.Accordingly, 242 annual and four two-year period difference layerswere obtained, exported from ENVI in Geotiff format, and subse-quently imported into an ArcGIS 10 environment (Fig. 3). Furtherprocessing was carried out by an ArcGIS spatial analyst in order todifferentiate changed from unchanged forest areas. A thresholdvalue for each difference change map was defined after analysis ofthe mean and standard deviation values of clear-cut areas.
3.4. Validation and rectification of results
The annual NDVI difference data (i.e. rough clear-cut areas) wereconverted from grid format into shapefile polygons. The borders ofthe shapes were then smoothed using the Smooth Polygon method
Fig. 2. Methodology for the detection of forest cover change.
P. Borrelli et al. / Applied Geography 48 (2014) 102e111 105
(PEAK smoothing algorithm,100-m tolerance) (Bodansky, Gribov, &Pilouk, 2002) to calculate the surface area of each polygon. Forestchange polygons that were smaller than 0.45 ha (less than 5Landsat pixels) were filtered out. Finally, each polygonwas buffered(5-m buffer) and resampled at 10-m cell size to bring the bound-aries of the algorithm-derived clear-cut areas more into line withthe real ones (Fig. 4).
The effective forest changes were displayed using semi-automatic mapping based on the multi-temporal set of annualforest cover change records. This procedure consisted of two steps:first, some multi-unsupervised forest cover classifications based onthe previous year’s Landsat images were carried out for the iden-tified forest change areas. Comparison of the resulting multi-temporal forest change values enabled many false-positivechanges to be identified and removed. Second, the remaining
Fig. 3. The 23 Landsat scenes covering Italian forest area.
false-positive forest changes were identified and filtered out bymeans of on-screen visual interpretation. More precisely, theannual polygon-vector data of forest change were overlapped withdigital aerial orthophotos with a geometric resolution of 0.25m (forthe years 2006 and 2011) on a large screen (24-inch) monitor. Thisvisual interpretation procedure allowed for the adjustment of theremaining false-positive clear-cuts.
Finally, through on-screen visual interpretation of different in-formation (e.g. Landsat RGB colour composition, annual NDVIchange detection rasters, validated forest cover change shapefiles),a subdivision was made between clear-cut areas and fire-affectedareas. Although visual interpretation is not fully effective indiscriminating these two forest disturbances, the observed firescars were found to be very different from the observed clear-cutareas. Indeed, as observed by other studies (Schroeder, Wulder,Healey, & Moisen, 2011, 2012; Chirici, Corona, Marchetti, Maselli,& Bottai, 2011), clear-cut areas appear with sharp borders and arather simple geometry, whereas fire scars are typically charac-terised by complex perimeters and fuzzy transitions between burntand unburnt areas.
3.5. Accuracy assessment
Accuracy assessment was carried out using the methodologyalready employed during the regional scale tests (Landsat image inPath 190, Row 31) (Borrelli et al., 2013a). The accuracy of the finalforest cover change product was archived through a confusionmatrix by a per-pixel analysis (Aronoff, 1982) and linear correlationanalysis (shapefile analysis). Accordingly, a set of the clear-cutareas, provided by an independent research group and alreadyused for the same purpose by Chirici et al. (2011), were used. Theauthors of this independent study mapped the forest harvesting byrecognition and delineation through multi-temporal on-screen vi-sual interpretation of aerial orthophotos and infrared false colourcomposite SPOT5 HRG images (study areae 25,100 km2 for the year2007). According to the methods applied for the on-screen visualinterpretation (for more info see Chirici et al., 2011), a minimummapping unit of 3 � 3 SPOT5 pixels was considered for clear-cutareas. Accordingly, only clear-cut areas within forest stands equalto or larger than 0.5 ha were considered, which fits well the spatialscale of our study. In addition, the database, the accuracy of whichwas found to be satisfactory through GPS field survey validation, isalso made up of only clear-cut area without other forest distur-bances (i.e. fires).
Fig. 4. Example of clear-cut detection and post-processing (Tuscany: 201251E 4711995N UTMWGS84-33). a) Landsat TM 432 false colour composite image (Path 192 Row 30, 2011);b) NDVI difference between Landsat TM 2010 and Landsat TM 2011; c) Original Landsat detected clear-cut (shapefile 30 � 30 m; d) Post-resampling clear-cut after PEAK smoothingalgorithm (100-m tolerance) and 5-m buffer.
P. Borrelli et al. / Applied Geography 48 (2014) 102e111106
4. Results
4.1. Forest cover change
The Landsat forest cover change map for Italy (Fig. 5) showschanges in forestland surface during the period from June 2002 toAugust 2011 (785.6 � 104 ha for 10 years ¼ 7856 � 104 ha). Theoverall results indicate that there were 125,272 individual clear-cutareas, corresponding to a total surface area of 317,535 ha. The sur-face of each clear-cut area detected through the Landsat NDVI dif-ferencing technique averages at 2.5 ha (s 4 ha). Of the total analysedsurface area of 785.6 � 104 ha, about 4.04% has been subject tosome form of wood extraction activities. From 2002 to 2006, thetotal change in forest area is estimated to be 160,501 ha, whereasfrom 2007 to 2011 it was about 157,034 ha. The annual rates offorest harvesting range from 0.3% (2008) to 0.5% (2002). Thegreatest harvesting of woodland occurred during 2002, affecting aforest area of 36,933 ha (Fig. 6).
At regional level (NUTS-2), Umbria, Lazio and Tuscany were theregions most exploited for wood extraction, accounting for 11.2%,10.5% and 8.2% respectively. By contrast, Lombardy, Liguria andAbruzzi were found to have the lowest number of clear-cut areas,accounting for 1.9% 1.7% and 1.9%, respectively (despite the fact thatthey are densely wooded). Very little coppice tree harvesting ac-tivity was found in Valle d’Aosta, Friuli-Venezia Giulia andTrentino-South Tyrol.
Fig. 7 illustrates the change in forest area at the EU NUTS-3 unit(province) level, while Fig. 8 shows examples of forest logging ac-tivities. On the provincial scale, the provinces with the most clear
cuts were Cremona (2193 ha; 43.6% of forestland), Lodi (1437 ha;43.1% of forestland), and Mantova (3853 ha; 41.2% of forestland).The largest forest change occurred in Perugia (23,763 ha; 11% offorestland).
Broad-leaved forest was the most harvested forest type(293,171 ha e 92.3%) followed by coniferous forest 16,385 ha e
5.2%) and mixed forest (7979 ha e 2.5%).Furthermore, some of the forest cover changes were found to
have occurred within areas declared as Sites of CommunityImportance (SCI) and Special Protection Areas (SPAs), as well asnational and regional parks such as Comprensorio Tolfetano-Cerite-Manziate, Montagnola Senese, Monti Simbruini and Ernici, andParco Nazionale Gran Sasso Monti della Laga (Table 2). Forestchange was detected in about 93,539 ha (equal to 29.5%) of SPAs.The protected area most severely affected was the ComprensorioTolfetano-Cerite-Manziate, where 5886 ha of forest area wereharvested over 10 years. In addition, about 58,264 ha of SCIs (18.3%of mapped forest cover changes) were mapped as having beensubjected to human-induced forest change.
The forest monitoring activity found Italian forests to be in goodhealth. Forest fires represent the major potential source of mappingerrors. A database dedicated to forest cover changes was compiledwith 4492 fire areas, corresponding to a total area of 61,503 ha.
4.2. Accuracy assessment
The visual interpretation of the aerial orthophotos helpedidentify actual forest cover changes. This procedure helped excludefalse indications of forest cover changes such as those caused by
Fig. 5. Forest change from July 2002 to July 2011. Note that the forest changes (brown)may be enhanced in size because of the small scale of the map. [For interpretation ofcolour referred in this figure legend, the reader is referred toweb version of the article.]
Fig. 7. Forest change area at province level (EU NUTS-3 units).
P. Borrelli et al. / Applied Geography 48 (2014) 102e111 107
seasonal variations in farmlands, pasturelands and rangelands,which are often erroneously included in the Corine land coverclassification (Büttner & Maucha, 2006; Büttner & Kosztra, 2007).
Both thematic and geometric accuracy assessments were per-formed, comparing clear-cut areas identified by the Landsat NDVIdifferencing techniquewith the clear-cut areas identified via an on-screen visual interpretation of aerial orthophotos and SPOT5 high-resolution geometry (HRG) images (years 2007 were provided byan independent research group). The thematic accuracy analysisshowed that 75.5% (n 1140; x 1.5 ha; s 2 ha) of visually identifiedclear-cut areas (larger than 0.45 ha) were also present in ouralgorithm-derived clear-cut database. By contrast, 369 clear-cutareas (x 1.4 ha; s 1.77 ha) detected by visual interpretation werenot detected by the algorithmic study. Eighty-six of these unde-tected clear-cut areas were found to fall into Landsat ETMþ (SLC-Off) gaps, so the actual number of undetected clear-cuts dropped to
Fig. 6. Forest cover change statistics (dark grey ¼ harvested area; lightgrey ¼ wildfires; black circles ¼ number of observed clear-cuts).
285. The linear correlation derived by comparing the clear-cut areasidentified via visual interpretationwith the clear-cut areas detectedby the Landsat NDVI differencing technique (Fig. 9; number ofpoints: 1140) results in a coefficient of determination (r2) of 0.96(a < 0.01). On the other hand, the Landsat NDVI results show thatan additional 1017 clear-cut areas (1145.5 ha; x 1.4 ha; s 1.9 ha),which were present in the algorithmic study database, were notpresent in the external validation database (Chirici et al., 2011). Thegeometric calculations were carried out on a per-pixel basis for the1140 clear-cut areas detected by both studies. The overall classifi-cation accuracy of the algorithm-derived clear-cut areas was 0.997,with a Kappa Index of Agreement (KIA) of 0.77.
5. Discussion
This article describes the methodology employed for thespatiotemporal mapping of forest cover change due to harvesting inItaly. It provides a consistent and spatially precise indication ofannual forest cover changes at the national scale for a significanttime period (2002e2011). In contrast to approaches based exclu-sively on fully automated monitoring operations e which are verytime-efficient but may be lacking in accuracy e the methodologyproposed allows for the mapping of only the actual forest coverchanges, based on thorough validation and rectification of thedetected forest perturbations. With a minimum mapping unit of0.45 ha (5 Landsat pixels), the study outcomes are presented in linewith international forestrymonitoring standards (Vidal et al., 2008).
The image differencing technique described in this study, basedon the Landsat NDVI, is shown to performwell in appropriately andaccurately monitoring forest harvesting in Mediterranean Europe.This is validated by procedures carried out in this study and in theprevious pilot project phase (Borrelli et al., 2013a). The feasibility ofthis study encourages further monitoring of forestry activities atthe European level, as have already been carried out for South
Fig. 8. Representation of the forest cover change detected (202e2011) in three different Italian locations highly involved in the wood supply chain: a) Grosseto (Tuscany) e 181216E4775713N UTM WGS84-33; b) Rieti (central Apennine) e 324700E 4686369 UTMWGS84-33; c) Cosenza (south Apennine) e 616092E 4335637N UTM WGS84-33. (Image database:Landsat).
P. Borrelli et al. / Applied Geography 48 (2014) 102e111108
America (Achard et al., 2002) Russia (Potapov et al., 2011) and Africa(Hansen et al., 2008). Moreover, the visual interpretation andrectification procedures based on aerial orthophotos allowed forthe selection of effective forest cover changes and the exclusion offalse indications that are often present in automatic mappingprocesses (Roy & Boschetti, 2009). The main observed sources offalse forest cover changes were: (i) the presence of agriculturalfields in the peripheral forest areas which showed greennessvariation related to the seasonality of crops, and (ii) pasturelands,which show great differences in NDVI values according to thevariability of the intensity of their yearly use. The on-screen visualrectification procedure enabled the removal of false forest changesthat were not identified during the previous automatic identifica-tion phase. While reducing the time-efficiency of the mappingprocess as a whole, this human-based procedure ensures thethorough validation and rectification of the outcomes. In addition,in order to minimise the presence of false positives to the greatestextent possible, this procedure was repeated using three differentspatial scales (i.e. 1:20,000, 1:10,000, and 1:5000).
An on-screen visual interpretation of orthophotos, Landsat RGB-band composition images and NDVI image difference grids also
revealed that the vast majority of the forest change areas detectedare made up of coppice forest clear cuts. During this phase, alldetected fires were filtered out and inserted into an independentdatabase. The results of the forest harvesting detection processshow that the clear-cut areas are of relatively small sizes (x 2.5 ha; s4 ha). Clear-cut areas of between 0.5 and 1.5 ha represent 20.2% ofthe total harvested forest area detected. The majority of these smallclear-cut areas were recorded by the approach adopted in thisstudy. In many cases, however, the shape of these small clear-cutareas cannot be detected using Landsat images because of therather coarse resolution of the sensors (for such small targets).Although it is possible to more effectively represent themorphology of the clear-cut areas that are larger than 1.5 ha, someinaccuracies were still noted. Most of the clear-cut areas that wereincorrectly classified as forest were located along the edges of forestand clear-cut areas. The occurrence of mixed pixels is most likely inthese areas, which represent a ‘mixture’ of the two land uses(Carroll, 2005). The clear-cut areas which were identified using on-screen visual interpretation, but were not mapped by our algo-rithmic detection procedure, correspond mostly to small forestlogging activities. In fact, 86.2% and 92.3% of clear-cut areas were
Table 2Forest cover change within protected areas.
Name of the natural reserve Forestchange
[ha] [%]
IT6030005 COMPRENSORIO TOLFETANO-CERITE-MANZIATE 5886 8.7IT8040021 PICENTINI 2843 4.5IT9210271 APPENNINO LUCANO, VALLE AGRI, MONTE
SIRINO, MONTE RAPARO2100 5.7
IT20B0501 VIADANA, PORTIOLO, SAN BENEDETTO POE OSTIGLIA
2029 28.1
IT6050008 MONTI SIMBRUINI ED ERNICI 1823 3.5IT6030043 MONTI LEPINI 1692 3.6IT5190003 MONTAGNOLA SENESE 1527 11.1IT5220008 MONTI AMERINI 1461 18.6IT8030008 DORSALE DEI MONTI LATTARI 1446 9.9IT6040043 MONTI AUSONI E AURUNCI 1444 2.3IT9130005 MURGIA DI SUD - EST 1350 2.8IT7110128 PARCO NAZIONALE GRAN SASSO - MONTI
DELLA LAGA1314 0.9
IT9310303 POLLINO E ORSOMARSO 1211 1.3IT8050052 MONTI DI EBOLI, MONTE POLVERACCHIO,
MONTE BOSCHETIELLO1166 8.1
IT6030085 COMPRENSORIO BRACCIANO-MARTIGNANO 1065 5.4IT6020005 MONTI REATINI 1005 4.3IT6030017 MASCHIO DELL’ARTEMISIO 993 39.6IT8050033 MONTI ALBURNI 945 4.0IT4010018 FIUME PO DA RIO BORIACCO A BOSCO
OSPIZIO915 14.9
IT51A0003 VAL DI FARMA 883 10.2IT8010006 CATENA DI MONTE MAGGIORE 875 16.9IT8040006 DORSALE DEI MONTI DEL PARTENIO 874 5.6IT8040013 MONTI DI LAURO 844 12.0IT8050055 ALBURNI 808 3.2IT9350300 COSTA VIOLA 786 4.2IT2080301 BOSCHI DEL TICINO 776 3.8IT2080501 RISAIE DELLA LOMELLINA 771 2.5IT6020017 MONTE TANCIA E MONTE PIZZUTO 743 10.9IT5170002 SELVA PISANA 735 7.7IT51A0017 CONO VULCANICO DEL MONTE AMIATA 728 11.9ITA030043 MONTI NEBRODI 701 1.0IT5190002 MONTI DEL CHIANTI 693 8.7IT5310018 SERRE DEL BURANO 644 17.7IT8050022 MONTAGNE DI CASALBUONO 640 3.7IT51A0029 BOSCHI DELLE COLLINE DI CAPALBIO 639 10.6IT9210275 MASSICCIO DEL MONTE POLLINO E MONTE ALPI 594 0.7IT7212134 BOSCO DI COLLEMELUCCIO - SELVAPIANA -
CASTIGLIONE594 9.5
IT51A0009 MONTE LEONI 576 11.3IT5190006 ALTA VAL DI MERSE 566 6.0IT9320302 Marchesato E FIUME NETO 561 0.8
P. Borrelli et al. / Applied Geography 48 (2014) 102e111 109
correctly identified by the algorithmic procedure for areas largerthan 1 ha and 3 ha respectively. The failure to identify smaller areasis due to a stronger effect of the aforementioned ‘mixture’ of thetwo land uses in a pixel (Carroll, 2005), which can be further
Fig. 9. Visual interpreted vs. algorithm-derived forest change.
exacerbated by the regeneration of grass and tree layers. Theoptimal conditions for identifying clear-cut areas are present (i)shortly after harvesting (Healey et al., 2005) and (ii) when rates ofgrass and tree layer regeneration are low (Wulder, Skakun, Kurz, &White, 2004). Given that 1017 clear-cut areas were detected in thisstudy using the algorithmic procedure but were not identified byvisual interpretation, wemay assume that our algorithm procedurefor identifying forest cover change is more effective that thehuman-based visual interpretation.
Moreover, the data analysis revealed that the average gaps in theLandsat ETM þ data (SLC-Off) amount, on average, to about 25% ofeach scene. The Landsat 7 ETM þ SLC-Off data are all Landsat 7images collected after May 31, 2003, when the Scan Line Corrector(SLC) failed (Maxwell, Schmidt, & Storey, 2007). To overcome thisissue, we undertook intensive mosaicing operations using theoverlapping surface of neighbouring images or extra images foreach year. This resulted in a remaining unmapped surface of about5% (2% due to ETMþ gaps and 3% due to cloud cover). We calculatedthat 86 of the 369 clear-cut areas that are missing from our data-base fall into a region that suffered from a bad combination ofLandsat ETMþ gaps and cloud cover. Had these clear-cut areas beendetected under normal Landsat ETMþ conditions (SLC-On), weassume that the thematic accuracy would have increased from75.5% to 81.2%.
The following reflections on the environmental impacts of theItalian forestry system are drawn from the observation of thespatial distribution and density of the clear-cut areas:
(i) From the spatial distribution of the clear-cut areas, it isevident that human-induced forest cover change occurs in allthe NUTS-2 level Italian units, at different levels of intensity.The peninsular regions that show the least forest cover dueto logging are the Veneto, Lombardy, Liguria Piedmont,Abruzzi and Apulia regions. However, the observed forestcover ranges from only 5.7% in Apulia to 21.6% in Veneto and62.6% in Liguria. This may reflect the fact that some regionalpolicies are more restrictive than others with respect towood extraction (e.g. Umbria, Lazio, Tuscany and Molise).
(ii) It was found that the Italian forests most involved in thewood supply chain are located in the mountains and hillyareas of the Italian peninsula. About 35% of the clear-cutareas are located on land characterised by moderate-to-steep slopes (>20�) with high precipitation values, whichare triggering factors for landslide occurrence (Imaizumi,Sidle, & Kamei, 2008) and soil degradation through erosion(Borrelli & Schütt, 2014; Borrelli, Märker, & Schütt, 2013b;Sorriso-Valvo, Bryan, Yair, Iovino, & Antronico, 1995; Stott,Leeks, Marks, & Sawyer, 2001).
(iii) There is an absence of compliance with the rules regardinglogging in Italy, as logging often takes place outside of thetime window allowed by law (from November to April). Inseveral cases reviewed as part of this study, trees were foundto be harvested during the summer period, with the furtherlack of post-harvest soil conservation measures.
(iv) Logging activities frequently occur within natural protectedareas. The harvested areas detected by this study includeseveral areas that are classified as Sites of CommunityImportance (SCI) and Special Protection Areas (SPAs). About93,539 ha of forests are harvested in areas subject to naturalprotection, with more than 50% of the total protected areabeing harvested in three areas (Sughereta di Tuscania, BoscoRonchetti and Bosco Selvapiana di Amaseno).
(v) For some provinces, the algorithm-derived clear-cut areameasurements tend to be higher than the officially reportedstatistics. In this studywemonitored the primary forest areas
P. Borrelli et al. / Applied Geography 48 (2014) 102e111110
and did not map ‘other wooded lands’ (CORINE scle-rophyllous vegetation e unit 3.2.3, transitional wood andscrub e unit 3.2.4). If the clear-cut areas of ‘other woodedlands’ were also mapped, the already significant differencedetected between the declared and actually harvested areas,which is consistent with the findings of other studies(Corona et al., 2007; Giuliarelli, 2009), may be even greater.
(vi) Most of the clear-cut areas typically range from1 to3ha in sizeas observed in past studies (Chirici et al., 2011; Corona et al.,2007; Giuliarelli, 2009). Small clear-cut areas are generallypreferred because, in most Italian administrative regions, theonly requirement for harvesting coppice areas that are lessthan 3 ha is that the forest ownermust submit a declaration tothe Italian State Forest Service (Corpo Forestale dello Stato).Accordingly, about 79% of the clear-cut areas detected in thisstudy are smaller than 3 ha and may have been clearedwithout the supervision of the responsible authorities.
(vii) Possible future increase of domestic forest resources. Ac-cording to Eurostat (2011), a relatively steady rise wasobserved in the level of roundwood production in the EU-27between 1995 and 2010, both for coniferous and non-coniferous species. In the near future, EU countries mayexperience an increase of domestic resource exploitation dueto the amendment of the European law banning illegallyharvested timber from the European market (FLEGT, 2011).
6. Conclusions
Wood harvesting is widely viewed as being a potentially serioushazard to forest health and a first-order triggering factor of envi-ronmental perturbation (e.g. soil degradation due to erosion, land-slides, drainage network pollution, dam siltation). Remote sensing-based approaches to mapping and monitoring spatiotemporalhuman-induced changes of forest cover havebeen the focus ofmuchresearch over the past decades. To date, however, most of theresearch in Italy has focused on the use of high-spatial-resolutiondata derived from sensors such as QuickBird (2.4-m), Ikonos (4-m)and SPOT5 (10-m), the cost of which is prohibitive for large-scaleassessments. The current study presents a novel method that usesLandsat imagery to assess large-scale forest-cover changes in spaceand time, and which also minimises false-positive changes. Theresults showed that forest clear-cut areas can bemappedwith goodaccuracy and for a sufficient time period, enabling further studies insupport of international biodiversity conventions. Even though thespatial resolution of the Landsat sensor imposes some limitations onthe detection of small logging areas, the proposed methodologycaptures the geographical distribution of human-induced forestchanges, and could facilitate the compilation of statistics regardingthe main sectors of wood extraction.
The final outcome, a comprehensive database of annual forestlogging activities in ESRI shapefile format, could be used in a vastnumber of research studies in various disciplines such as biology,geography, soil science, engineering and economics. It is currentlybeing further processed to precisely identify the time period of eachclear-cutting activity. This can be achieved with the current meth-odology on a sub-annual scale by using several Landsat scenes foreach year (e.g. for the months of April, June, August and October).
Acknowledgements
The authors are grateful to National Aeronautics and SpaceAdministration (NASA) and United States Geological Survey (USGS)who have donated the Landsat imagery. The authors also wish tothank Gráinne Mulhern for English-language editing.
References
Achard, F., Eva, H. D., Stibig, H. J., Mayaux, P., Gallego, J., Richards, T., et al. (2002).Determination of deforestation rates of the world’s humid tropical forests.Science, 297, 999e1002.
APAT (2005). Carta Geologica d’Italia in scala 1: 1.250.000. [Online] URL. http://www.isprambiente.gov.it/images/progetti/progetto-1250-ita.jpg. Accessed July2011
Aronoff, S. (1982). Classification accuracy: a user approach. Photogrammetric Engi-neering & Remote Sensing, 48, 1299e1307.
Asner, G. P., Keller, M., Pereira, R., Jr., & Zweede, J. C. (2002). Remote sensing ofselective logging in Amazonia: assessing limitations based on detailed fieldobservations, Landsat ETMþ, and textural analysis. Remote Sensing of Environ-ment, 80, 483e496.
Bodansky, E., Gribov, A., & Pilouk, M. (2002). Smoothing and compression of linesobtained by raster-to-vector conversion. In D. Blostein, & Young-Bin Kwon(Eds.), Graphics recognition algorithms and applications (pp. 256e265). SpringerBerlin Heidelberg.
Borrelli, P., Sandia Rondón, L. A., & Schütt, B. (2013a). The use of Landsat imagery toassess large-scale forest cover changes in space and time, minimizing false-positive changes. Applied Geography, 41, 147e157.
Borrelli, P., Märker, M., & Schütt, B. (2013b). Modelling post-tree-harvesting soilerosion and sediment deposition potential in the Turano River Basin (ItalianCentral Apennine). Land Degradation & Development. http://dx.doi.org/10.1002/ldr.2214. in press.
Borrelli, P., & Schütt, B. (2014). Assessment of soil erosion sensitivity and post-timber-harvesting erosion response in a mountain environment of Central Italy.Geomorphology, 204, 412e424.
Büttner, G., & Kosztra, B. (2007). CLC2006 technical guidelines. EEA. Technical report17/2007, ISSN 1725-2237.
Büttner, G., & Maucha, G. (2006). In The thematic accuracy of Corine land cover 2000,assessment using LUCAS (land use/cover area frame statistical survey). Copenha-gen: European Environment Agency. Technical report/no7/2006.
Carroll, M. (2005). Rapid assessment of annual deforestation in the BrazilianAmazon using MODIS data. Earth Interactions, 9, 1e22.
Chavez, P. S., & Mackinnon, D. J., Jr. (1994). Automatic detection of vegetationchanges in the southwestern United States using remotely sensing images.Photogrammetry and Remote Sensing, 60, 571e583.
Chen, J. M., & Cihlar, J. (1996). Retrieving leaf area index of boreal conifer forestsusing Landsat TM images. Remote Sensing of Environment, 55, 153e162.
Chirici, G., Corona, P., Marchetti, M., Maselli, F., & Bottai, L. (2011). In A. Amaro,D. Reed, & P. Soares (Eds.), Spatial distribution modelling of forest attributescoupling remotely sensed imagery and CIS techniques (pp. 41e50). Oxford:Modeling Forest Systems.
Cochrane, M. A. (2003). Fire science for rainforests. Nature, 421, 913e919.Cohen, W. B., Spies, T. A., Alig, R. J., Oetter, D. R., Maiersperger, T. K., & Fiorella, M.
(2002). Characterizing 23 years (1972e95) of stand replacement disturbance inwestern Oregon forests with Landsat imagery. Ecosystems, 5, 122e137.
Collins, J. B., & Woodcock, C. E. (1996). An assessment of several linear changedetection techniques for mapping forest mortality using multitemporal landsatTM data. Remote Sensing of Environment, 56, 66e77.
Coops, N. C., Gillanders, S. N., Wulder, M. A., Gergel, S. E., Nelson, T., & Goodwin, N. R.(2010). Assessing changes in forest fragmentation following infestation usingtime series Landsat imagery. Forest Ecology and Management, 259, 2355e2365.
Coppin, P. R., & Bauer, M. E. (1994). Processing of multitemporal Landsat TM im-agery to optimize extraction of forest cover change features. Geoscience andRemote Sensing, 32, 918e927.
Coppin, P. R., & Bauer, M. E. (1996). Digital change detection in forest ecosystemswith remote sensing imagery. Remote Sensing Reviews, 13, 207e234.
Corona, P., Giuliarelli, D., Lamonica, A., Mattioli, W., Tonti, D., & Chirici, I. G. (2007).Confronto sperimentale tra superfici a ceduo tagliate a raso osservate medianteimmagini satellitari ad alta risoluzione e tagliate riscontrate amminis-trativamente. Forest, 4, 324e332.
EEA, European Environmental Agency. (2011). [Online] URL. http://www.eea.europa.eu/data-and-map Accessed July 2011.
Elmore, A. J., Mustard, J. F., Manning, S. J., & Lobell, D. B. (2000). Quantifyingvegetation change in semiarid environments: precision and accuracy of spectralmixture analysis and the normalized difference vegetation index. RemoteSensing of Environment, 73, 87e102.
European Commission. (2003). Sustainable forestry and the European Union e ini-tiatives of the European Commission. Office for Official Publications of the Eu-ropean Communities.
European Commission. (2006). Thematic strategy for soil protection. COM, 231.European Commission. (2011). Our life insurance, our natural capital: an EU
biodiversity strategy to 2020. COM, 244. Luxembourg.Eurostat. (2011). Forestry in the EU and the world. Statistical books. Luxembourg:
Publications Office of the European Union.FAO. (2010). Global forest resources assessment 2010. Rome: Italy.FAO. (2012). State of the World’s forests. Rome: Italy.FLEGT, Forest Law Enforcement, Governance and Trade. (2011). [online] URL. http://
www.euflegt.efi.int/portal/home/eu_timber_regulation/ Accessed December2011.
Giuliarelli, D. (2009).Utilizzodi immagini satellitari adalta risoluzioneper ilmonitoraggiodelle superfici tagliate a raso in boschi cedui. Dissertation. University of Tuschia.
P. Borrelli et al. / Applied Geography 48 (2014) 102e111 111
Hansen, A. J., Neilson, R. P., Dale, V. H., Flather, C. H., Iverson, L. R., Currie, D. J., et al.(2001). Global change in forests: responses of species, communities, and bi-omes. Bioscience, 51, 765e779.
Hansen, M. C., Roy, D. P., Lindquist, E., Adusei, B., Justice, C. O., & Altstatt, A. (2008).A method for integrating MODIS and Landsat data for systematic monitoring offorest cover and change in the Congo Basin. Remote Sensing of Environment, 112,2495e2513.
Healey, S. P., Cohen, W. B., Zhiqiang, Y., & Krankina, O. N. (2005). Comparison oftasseled cap-based Landsat data structures for use in forest disturbancedetection. Remote Sensing of Environment, 97, 301e310.
Huete, A., Didan, K., Miura, T., Rodriguez, E. P., Gao, X., & Ferreira, L. G. (2002).Overview of the radiometric and biophysical performance of the MODISvegetation indices. Remote Sensing of Environment, 83, 195e213.
Imaizumi, F., Sidle, R. C., & Kamei, R. (2008). Effects of forest harvesting on theoccurrence of landslides and debris flows in steep terrain of central Japan. EarthSurface Processes and Landforms, 33, 827e840.
INFC. (2005). Linee generali del progetto per il secondo inventario forestale nazionaleitaliano. Trento: Inventario Nazionale delle Foreste e dei Serbatoi Forestali diCarbonio, MiPAF e Ispettorato Generale del Corpo Forestale dello Stato, CRA-ISAFA.
INFC. (2007). Le stime di superficie 2005. Trento: Ministero delle Politiche Agricole,Alimentari e Forestali, Ispettorato Generale e Corpo Forestale dello Stato, CRA e
Istituto Sperimentale per l’Assestamento Forestale e l’Alpicoltura.Jensen, J. R. (1986). Introductory digital image processing. New Jersey: Prentice-Hall.Kaplan, J. O., Krumhardt, K. M., & Zimmermann, N. (2009). The prehistoric and
preindustrial deforestation of Europe. Quaternary Science Reviews, 28, 3016e3034.
Kennedy, R. S., & Spies, T. A. (2004). Forest cover changes in the Oregon coast rangefrom 1939 to 1993. Forest Ecology and Management, 200, 129e147.
Kyoto Protocol. (1997). United Nations framework convention on climate change.Kyoto: Kyoto Protocol.
Li, M., Huang, C., Zhu, Z., Shi, H., Lu, H., & Peng, S. (2009). Assessing rates of forestchange and fragmentation in Alabama, USA, using the vegetation changetracker model. Forest Ecology and Management, 257, 1480e1488.
Linke, J., McDermid, G. J., Pape, A., McLane, A. J., Laskin, D. N., Hall-Beyer, M., et al.(2009). The influence of patch delineation mismatches on multi-temporallandscape pattern analysis. Landscape Ecology, 24, 157e170.
Liu, H. Q., & Huete, A. (1995). A feedback based modification of the NDVI to mini-mize canopy background and atmospheric noise. Geoscience and RemoteSensing, 33, 457e465.
Lu, D., Mausel, P., Brondízio, E., & Moran, E. (2004). Change detection techniques.International Journal of Remote Sensing, 25, 2365e2401.
Lyon, J. G., Yuan, D., Lunetta, R. S., & Elvidge, C. D. (1998). A change detectionexperiment using vegetation indices. Photogrammetric Engineering and RemoteSensing, 64, 143e150.
Mas, J. F. (1999). Monitoring land-cover changes: a comparison of change detectiontechniques. International Journal of Remote Sensing, 20, 139e152.
Maxwell, S. K., Schmidt, G. L., & Storey, J. C. (2007). A multi-scale segmentationapproach to filling gaps in Landsat ETMþ SLC-off images. International Journal ofRemote Sensing, 28, 5339e5356.
McFeeters, S. K. (1996). The use of normalized difference water index (NDWI) in thedelineation of open water features. International Journal of Remote Sensing, 17,1425e1432.
MCPFE. (2009). Ministerial conference on the protection of forests in Europe [online]URL http://www.mcpfe.org Accessed May 2010.
Panagos, P., Van Liedekerke, M., Jones, A., & Montanarella, L. (2012). European SoilData Centre (ESDAC): response to European policy support and public datarequirements. Land Use Policy, 29, 329e338.
Pompei, E., & Gasparini, P. (2007). Compiti e finalità dell’Inventario delle foreste.Silvae, 9, 1e26.
Potapov, P., Hansen, M. C., Stehman, S. V., Loveland, T. R., & Pittman, K. (2008).Combining MODIS and Landsat imagery to estimate and map boreal forestcover loss. Remote Sensing of Environment, 112, 3708e3719.
Potapov, P., Turubanova, S., & Hansen, M. C. (2011). Regional-scale boreal forestcover and change mapping using Landsat data composites for European Russia.Remote Sensing of Environment, 115, 548e561.
Richards, J. F., & Tucker, R. P. (1988). World deforestation in the twentieth century.Durham: Duke University Press.
Rodriguez-Galiano, V., & Chica-Olmo, M. (2012). Land cover change analysis of aMediterranean area in Spain using different sources of data: multi-seasonalLandsat images, land surface temperature, digital terrain models and texture.Applied Geography, 35, 208e218.
Roy, D. P., & Boschetti, L. (2009). Southern Africa validation of the MODIS, L3JRC, andGlobCarbonburned-areaproducts.GeoscienceandRemoteSensing, 47,1032e1044.
Rusco, E., Filippi, N., Marchetti, M., & Montanarella, L. (2003). Carta Ecopedologicad’Italia. Scala 1: 250.000. Technical EUR reports. Ispra, Italy: European Soil Bureau.
Schroeder, T. A., Wulder, M. A., Healey, S. P., & Moisen, G. G. (2011). Mapping wildfireand clearcut harvest disturbances in boreal forests with Landsat time seriesdata. Remote Sensing of Environment, 115, 1421e1433.
Schroeder, T. A., Wulder, M. A., Healey, S. P., & Moisen, G. G. (2012). Detecting post-fire salvage logging from Landsat change maps and national fire survey data.Remote Sensing of Environment, 122, 166e174.
Singh, A. (1989). Digital change detection techniques using remotely-sensed data.International Journal of Remote Sensing, 10, 989e1003.
Sorriso-Valvo, M., Bryan, R. B., Yair, A., Iovino, F., & Antronico, L. (1995). Impact ofafforestation on hydrological response and sediment production in a smallCalabrian catchment. Catena, 25, 89e104.
Stott, T., Leeks, G., Marks, S., & Sawyer, A. (2001). Environmentally sensitive plots-cale timber harvesting: impact on suspended sediment. Bed load and bankerosion dynamics. Journal of Environmental Management, 63, 3e25.
UNFCCC. (2007). Investment and financial flows to address climate change. Bonn,Germany: UNFCCC.
USGS. (2011). Landsat product type descriptions [online] URL: http://edcsns17.cr.usgs.gov/helpdocs/landsat/product_descriptions.html Accessed January 2011.
Vacchiano, G., Magnani, F., & Collalti, A. (2012). Modeling Italian forests: state of theart and future challenges. iForest, 5, 113e120.
Vidal, C., Lanz, A., Tomppo, E., Schadauer, K., Gschwantner, T., di Cosmo, L., et al.(2008). Establishing forest inventory reference definitions for forest andgrowing stock: a study towards common reporting. Silva Fennica, 42, 247e266.
Williams, M. (2000). Dark ages and dark areas: global deforestation in the deeppast. Journal of Historical Geography, 26, 28e46.
Williams, D. (2006). Landsat-7 science data users handbook [online] URL http://ltpwww.gsfc.nasa.gov/IAS/handbook/handbook_htmls/chapter11/chapter11.html Accessed November 2011.
Wilson, E. H., & Sader, S. A. (2002). Detection of forest harvest type using multipledates of Landsat TM imagery. Remote Sensing of Environment, 80, 385e396.
Woodcock, C. E., Allen, R., Anderson, M., Belward, A., Bindschadler, R., & Cohen, W.(2008). Free access to Landsat imagery. Science, 320, 1011.
Woodcock, C. E., Macomber, S. A., Pax-Lenney, M., & Cohen, W. B. (2001). Moni-toring large areas for forest change using Landsat: generalization across space,time and Landsat sensors. Remote Sensing of Environment, 78, 194e203.
Wulder, M. A., Skakun, R. S., Kurz, W. A., & White, J. C. (2004). Estimating time sinceforest harvest using segmented Landsat ETMþ imagery. Remote Sensing ofEnvironment, 93, 179e187.
Zha, Y., Gao, J., & Ni, S. (2003). Use of normalized difference built-up index inautomatically mapping urban areas from TM imagery. International Journal ofRemote Sensing, 24, 583e594.
Web references
Lunetta, R. S., Johnson, D. M., Lyon, J. G., & Crotwell, J. (2004). Impacts of imagerytemporal frequency on land-cover change detection and monitoring. RemoteSensing of Environment, 89, 444e454.
Sexton, J. O., Urban, D. L., Donohue, M. J., & Song, C. (2013). Long-term land coverdynamics by multi-temporal classification across the Landsat-5 record. RemoteSensing of Environment, 128, 246e258.