One of the most useful applications of remote sensing is extracting thematic information and recognize patterns from raster images using image classification techniques (Jensen, 2016). Due to the low-cost and large-area coverage forest information, satellite remote sensing has been considered best for mapping broad scale forestry, as there are various methods for identifying patterns and various for such as classification, statistical analysis etc. (Franklin et al., 2003). Different objects on the earth’s surface have a different spectral reflectance and remittance properties (Al-doski, Mansor, & Shafri, 2013). Based on this idea, image classification is defined as the process of categorizing all pixels in an image or raw remotely sensed satellite data to obtain a given set of labels or land cover classes or themes (Kiefer, Lillesand, & Chipman, 2015).
There has always been a major concern with the scale issue, in the field of remote sensing as the level of spatial scale can influence the image attraction process and output (Wu & Li, 2009). When image pixels are smaller than the objects of interest, information extraction techniques like classification can be applied (Department of Geography, 2017). However, when image pixels are larger than the objects of interest the user has to choose alternate approaches, like statistical analysis/ regression analysis (Department of Geography, 2017). This assignment discusses about the application of classification technique on Landsat 5 image, exploring the statistical relationship among different variables for predicting leaf area index (LAI) and crown closure, and creating composite maps by combining classified layer and continuous variable information. The overall objective of this assignment can be summarized as :
1. To perform a multi-scale information extraction of landcover and vegetation structure from Landsat TM imagery and DEM data, using classification and statistical modelling techniques
2. To evaluate and communicate the results of your analysis through accuracy statistics, construction of map products, and production of a formal lab report.
This assignment aimed at performing the supervised classification on Landsat 5 image based on the land cover data, creating two continuous variable maps of LAI index and crown closure, and creating two composite maps. In the supervised classification map, the overall accuracy was 60.667%, which was not high enough. The confusion matrix table showed that there were overlapping of land cover classes. Statistical analysis like regression analysis was useful in understanding which variable was a good predictor. The result obtained from the multiple regression analysis using six variables showed that the wetness variable had the highest coefficient of determination with LAI (R2 = 0.69) and crown closure (R2 = 0.52). However, while checking the assumption of the regression it was found out that the dependent variables (LAI and crown closure) were not normally distributed and the variance of the residuals were not homoscedastic. Hence, the result of the regression model cannot be entirely trusted. Two composite maps, produced by combining broadleaf forest land cover and LAI index and coniferous forest and crown closure, provided information regarding the distribution of LAI and crown closure ranges for the mentioned forest vegetation classes. For future studies, it is suggested to collect more ground truth data and use imagery of higher resolution. In addition, composites map can be created for multiple forest vegetation classes by combining those with Lai and crown closure. Different other indices like RSR, SAVI and MSAVI or slope and aspect could be also used in the regression model to see if those can make the model better, if time and resource permits.