1.0 Class Description

2.0 Map Description

The statewide land-use/land-cover map of Massachusetts was created by the Massachusetts Forest Monitoring Program (MaFoMP) employing remote sensing classification techniques utilizing Landsat TM/ETM+ imagery from 2000-2002 and environmental GIS variables.  Five Landsat scenes were classified independently and mosaicked together to cover the entire state of Massachusetts. Classification was performed using a pixel based Classification Tree (based on the See5 algorithm), followed by a 3x3 majority filter to produce realistic landscape management units.

Satellite data employed for classification of each scene included imagery captured in September (representing onset of vegetation senescence in Massachusetts) and October (representing advanced vegetation senescence in Massachusetts). Environmental GIS variables included in the classification were elevation, slope, precipitation, and surficial geology (used as a proxy for unavailable soil information).

3.0 Map Accuracy

Map accuracy was determined through independent validation ground reference data utilizing the Kappa Index of Agreement (KIA). Overall kappa accuracy of the statewide map was 0.82 with per-class kappa values ranging from 0.69 (grassland) to 0.96 (salt marsh). Accuracy of the independently classified scenes used to create statewide coverage ranged from 0.81 to 0.84.

Since the ultimate goal of the project is to identify forest cover across Massachusetts we took care to correctly identify forest locations. While there was a certain amount of class confusion between forest types (e.g., Deciduous confused with Mixed, and Conifer confused with Mixed) overall we saw a low error of omission (did not correctly classify forest in areas where it does exist) at 1.24%. This error can be mostly accounted for in class confusion between land types of similar properties such as low density developed, where the landscape is a mix of forested and impervious surface, orchards and wetlands (see 4.0 class confusion).

4.0 Class Confusion

A seventeen category map legend an ambitious level of detail for Landsat imagery (30m) and as a result many of our classes were similar to each other in reflectance properties lending to varying degrees of interclass confusion. In reviewing the trouble areas it is evident that the majority of inter-class confusion occurs between land-use categories of similar ground cover. Confusion in the forest classes occurred between mixed forest and deciduous forest, as well as mixed forest with conifer forest, however pure conifer and deciduous forests were separable from each other.  The bare land classes characterized by sparse or lack of vegetation were primarily mistaken with each other, the residential classes and commercial/ industrial confused, as well as sand quarry and the aforementioned developed classes. Herbaceous dominant land-use categories such as golf courses, grasslands, and pasture all confused with each other. Orchards mixed more often with the herbaceous classes than forested, which can be attributed to the greater proportion of well managed grass cover to sparsely separated deciduous canopies characteristic of commercial fruit plantations in the state. Classes such as water and salt marsh which have unique electromagnetic reflectance characteristics from all other classes saw little to no interclass confusion. 

To asses the accuracy with which the forest cover was correctly identified beyond the map statistics, a comparison with a different land cover product was performed. The MassGIS land-use map layer contains 37 land use categories that were interpreted from 1:25,000 aerial photographs taken in 1999 with a minimum mapping unit of approximately 0.4 ha (http://www.mass.gov/mgis/lus.htm). The MassGIS land-use layer contains only one forest category, so for the purpose of comparison Deciduous, Conifer, and Mixed classes from the MaFoMP classification were merged to identify forest cover only. Simple map algebra techniques (i.e. image differencing) were used to reveal locations of discrepancies between the two maps. A qualitative analysis indicated that these differences were a result of: 1) different map legends; or 2) the difficultly in accurately distinguish the low residential class with medium spatial resolution imagery.

Some examples of these discrepancies follow.

MassGIS only

In the central-western portion of the state, along the Connecticut River, areas that were identified as forested in the MassGIS layer, were classified as wetlands by the MaFoMP. For a further investigation a wetland layer was acquired. The DEP Wetland datalayer was produced by the Department of Environmental Protection (DEP) in collaboration with the Univeristy of Massachussetts, Amherst (UMASS, Amherst) and is distributed through MassGIS.  The layer contains 28 types of wetlands that were interpreted from 1:12,000 stereo color-infrared (CIR) photography (http://www.mass.gov/mgis/wetdep.htm). When the MaFoMP land cover and the DEP Wetlands layer were overlaid, a lot of agreement was noticed.  Aerial photo interpretation (on Google Earth) confirmed that the areas identified as forest only by MassGIS were actually wetlands.  Therefore, we believe that these areas were correctly identified as wetlands in the MaFoMP land cover map.

MaFoMP only

According to the forest comparison the low density, peri-urban areas west of Boston were incorrectly classified as forested land. MassGIS identified these areas as Residential 2 and Residential 3. Investigation of this problem revealed that these discrepancies were a result of different land use code definitions that were used for the two products. For instance, Residential 3 is defined by MassGIS as residential lots larger than ½ acre. This class corresponds roughly to an area where there is one house every 2000+ m². Given the spatial resolution of 30 m that was used to produce the MaFoMP land cover, 2000 m² is equivalent to a little over two 30 m pixels. Therefore, it is reasonable to believe that while one pixel might have been classified as low residential the other one was classified as a vegetated class, most likely forest. This analysis indicates that the minimum mapping unit (MMU) and class definition play a major role in the final classified output.

Another factor that can account for the differences in the two classifications is the difficultly in discriminating low-density residential areas using a medium spatial resolution of 30 m. A pixel might contain some residential features and a wide range of land cover types such as forest, grassland or a combination of both. The built areas have a different spectral response than the vegetated features but they all contribute in determining the value of that single pixel. Hence, the difficulty in accurately mapping this particular class.