Method used

Extraction of trees from multi resolution remote sensing data (outlined in Fig. 2), involved.

(i) Fusion of multi resolution data for optimal spatial and spectral resolutions. Data fusion was performed using various algorithm (such as hyperspectral color space resolution merge, high pass filter fusion, modified intensity hue saturation fusion, wavelet fusion) and performance is evaluated using UIQI index (equation (4)), which helps in measuring the similarity and distortion by considering three factors namely loss of correlation between datasets, radiometric imbalance, and contrast distortion. UIQI index was calculated using,

--------------------------------- (4)



Fig. 2. Method involved in extraction of tree cover of Bangalore.

(a)Data Collection :   Mapping the spatial extent of species wise tree canopy in select wards using the pre calibrated GPS and through high spatial resolution data including Google Earth. b) Land use:  The spatial extent of vegetation cover was derived from classified remote sensing data. Field data (involving number of trees and spatial extent of tree crown) of sampled wards was compared with the vegetation cover, which helped in getting the number of trees in each ward. c) Frequency distribution:   The spatial extent of vegetation cover was derived from classified remote sensing data. Field data (involving number of trees and spatial extent of tree crown) of sampled wards was compared with the vegetation cover, which helped in getting the number of trees in each ward. Based on the field data and virtual online database, sampled wards were grouped into (i)> 500 trees and (ii) < 500 trees. Frequency distribution of number of trees versus average area of tree canopy was analyzed - local maxima, local minima and also edges (with other sub classes) Knowledge of species wise local maxima and minima aided in understanding probable location ranges with the respective local peak in radiometric brightness (fused data in NIR band). In cases of near neighborhood of a species or group of tree species the maxima and minima is computed based on geometric mean of neighbors at 8 search directions. The choice of which maxima or minima are ultimately considered valid is based on the threshold. These would aid as sampling interval points for extracting pixels from classified land use data. d) Population projection for 2013:   The population for the year 2013 was estimated based on the decadal growth (equation (5)). P2013(i)=P2011(i)*(1 + n * r(i))------------------------------(5)

(ii) land use analyses to understand the land use dynamics e thisinvolved a) spatial data preprocessing (geo rectification using reference data) and generation of False Color Composite (FCC) of remote sensing data (bands e green, red and NIR). FCC helped in locating heterogeneous patches in the landscape, b) selection of training polygons (these correspond to heterogeneous patches in FCC) covering 15% of the study area and uniformly distributed over the entire study area, c) loading these training polygons co-ordinates into precalibrated GPS, d) collection of the corresponding attribute data (land use types) for these polygons from the field. GPS helped in locating respective training polygons in the field, e)supplementing this information with Google Earth, f) 60% of the training data has been used for classification, while the balance is used for validation or accuracy assessment. Land use analysis was carried out through supervised pattern classifier - Gaussian maximum likelihood algorithm using fused spatial data. The MLC classifier uses probabilities(Lillesand et al., 2015 to classify each pixel into a particular land use class (categories: Built up; Vegetation; Water; Others). The Gaussian MLC classification technique has been used widely for analysis of land use as this technique is proved to be more superior that other classification techniques (Duda et al., 2000).,
(iii) Statistical assessment of classifier performance based on the performance of spectral classification considering reference pixels is done which include computation of kappa (k) statistics and overall (producer's and user's) accuracies (Ramachandra et al., 2012a,b, 2015; Bharath and Ramachandra, 2016). For earlier time data, training polygon along with attribute details were compiled from the historical published topographic maps, vegetation maps, revenue maps, etc.
(iv) Field data collection from select wards-mapping of trees using pre calibrated GPS. This involved inventorying, mapping and measurement of tree canopy.
(v) Ward wise trees distribution analyses based on canopy size and spectral analyses of tree canopy (based on species and age). This involved
               Where
               P2013(i) - Population of ward i for the year 2013.
               P2011(i) - Population of ward i for the year 2011.
               n - Number of decades ¼ 0.2.
               r(i) e Incremental rate of ward i.
The ratio of number of trees in each ward to population in each ward (equation (6)) was determined to quantify tree distribution per person in each ward. Similarly trees per person in Bangalore city is computed (equation (7))


-------------------------(6)

-------------------------(7)

       Where
       TpP(i)- Tree per person in ward i.
       Tree(i) - Number of trees in ward i.
       TpP(B) - Tree per person in Bangalore

(vi) Validation: Validation of tree count for select ward is done by comparing with actual count of trees based on field work (equation (8)).

-------------------------(8)

                Where
               Class_Tree - count based on classified data.
               GPS_Tree- Tree count based on field census using GPS.