| Back |
This study highlights the various preprocssing techniques and the classification algorithms (hard and soft) used in the research and summarises the results obtained from the accuracy assessment. The research questions have been answered after an insight in to the findings of the result. Finally the chapter concludes with the scope for further research.
This work compares and assesses the efficacy of various hard classification algorithms namely K-Means Clustering, Maximum Likelihood Classifier (MLC), Spectral Angle Mapper (SAM), Neural Network (NN) and Decision Tree Approach (DTA) as well as Linear Mixture Model (LMM) as a soft classifier for land cover mapping using the MODIS (250 m spatial resolution) surface reflectance product.
Principal Component Analysis (PCA) and Minimum Noise Fraction (MNF) were used to remove the redundancy and noise from the MODIS data. To begin with, NDVI helped in obtaining the general land cover (green versus non-green) mapping for study area (Kolar district, Karnataka state, India) using LISS-3 (23.5 m spatial resolution) data and MODIS data. LISS-3 data were classified using supervised MLC into 6 broader land cover categories (Agriculture, Built up (Urban / Rural), Evergreen / Semi-Evergreen forest, Plantations / orchards, Waste lands / Barren rock/ Stony waste / Sheet rock and Waterbodies / Lakes / Ponds / Tanks / Wetland).
In the first phase of experiments, MODIS Bands 1 to 7, PCA-derived components and MNF components were classified into the same 6 land cover classes using K-Means clustering, Maximum Likelihood Classifier, Spectral Angle Mapper, Neural Network and Decision Tree Approach. In the second phase, 6 fraction images pertaining to each class were generated using the Linear Mixture Model.
The error matrix for LISS-3 classified map showed an overall accuracy of 95.63 % for Chikballapur taluk (table 6.2) where field data were collected. Accuracy assessment of the MODIS hard classified maps were estimated at three levels.
• Error matrices were generated with ground truth data for Chikballapur taluk. The overall accuracy, producer's and user's accuracy were computed. Neural Network based classification on MNF components showed the best overall accuracy of 86.11 % and MLC on MODIS bands 1 to 7 was the second in ranking with 75.99 % accuracy (see table 6.3 for details). MLC on Principal Components ranked 13 th among the techniques with an overall accuracy of 30.44 %.
• A comparison of the land cover percentage area among the various classification algorithms was done for Chikballapur taluk (table 6.6). The results showed that MLC on MODIS Bands 1 to 7 is best for mapping agriculture and built up. K-Means algorithm was best for mapping forest. Decision Tree Approach on PCs was found to be good for plantation, while the same technique on MODIS Bands 1 to 7 was best for mapping waste land. Neural Network on MNF component mapped the waterbodies accurately.
At pixel level, MODIS hard classified maps were compared to high resolution LISS-3 classified map. The corresponding producer's, user's and overall accuracy were computed. This assessment was in agreement with the error matrix computed using ground truth. It showed that
• Neural Network classification on MNF components has the highest accuracy of 69.87 % and MLC on MODIS Bands 1 to 7 had an accuracy of 65.77 % in decreasing order (see table 6.7 for details). The other rankings did not tally exactly with the rankings computed using ground truth and were shuffled to 2 or 3 positions high or low in the rankings. This may be due to the following reasons. While comparing the LISS-3 and MODIS classified maps, only 100% pure land cover classes (the matrix of 11 x 11 pixels in LISS-3 classified map) were considered and those ranging from 90 – 100 % were neglected. Also, 11 pixels of LISS-3 are 258.5 m spatially, which is 8.5 m more than the spatial resolution of MODIS pixel (250 m). To summarize, the above discussion shows the utility / usefulness of the existing hard classification techniques for land cover mapping at regional scale using MODIS data.
The abundance maps that could give the fraction of each class in a pixel were also verified at administrative boundary (taluk level) as well as at pixel levels.
• (i.) At the taluk level, the comparison was based on the percentage area of each class with respect to LISS-3 classified map and the percentage of the classes were found to be similar (see table 6.10 for details). The percentage of waterbodies was higher in the abundance map.
• (ii.) At the pixel level, the accuracy of the proportions of the classes was validated on the ground. There is mixing of classes in agriculture--plantation/orchards, forest--plantations and built up--waste land/rocky/stony/barren (see table 6.11). Waterbodies were overestimated in the study area, indicating requirement of adequate training sites / endmembers representing water category. MODIS land cover outputs is comparable to LISS-3 classified map, evident from the validation exercise, as most endmembers (pure pixels) were correctly classified while mixed pixels were within 20-25% of the actual values. The result shows the utility of spectral unmixing model to obtain the abundance maps of the different interpretable classes.
The consequence of different spatial scales (at administrative boundary level and at pixel level) on classification accuracy was noticed. When comparing the land cover percentage area among the various algorithms with respect to LISS-3 classified image, it was observed that for mapping a particular land cover class at the boundary level, say for agriculture and built up MLC on MODIS bands 1 to 7 was the best but when the accuracy was assessed at pixel level, NN on MNF components proved to be superior than the other techniques. A certain algorithm may be good for mapping a particular class but at the same time may not be equally good for mapping all other classes. Further, it is recommended to rely on pixel level result than at the boundary level. To interpret similar “percentages” from different classifiers as an indicator for correct classification is not advisable and reliable. The percentage area of class “A” for two classifiers may be very similar on an administrative boundary level (i.e. for the entire study region) but quite different if the classified images are compared pixel-by-pixel. Pixel analysis gives an idea about the “stability” of classification results.
The preprocessing techniques (Principal Component Analysis) and Minimum Noise Fraction (MNF) on MODIS data had different effects on the accuracy of the hard classification algorithms. It was observed that Neural Network (NN) and Decision Tree Approach (DTA) on MNF components give the best results as is evident from their ranking (see table 6.3 for details) which was generated on the basis of overall accuracy from ground truth. NN and DTA approach on PC's also gave good result and ranked 5 th and 6 th among the algorithms. The interesting point here is Spectral Angle Mapper and Maximum Likelihood classification that performed worst on the MNF components as well as PCs. The same observation was made in the accuracy assessment based on pixel to pixel comparison with respect to LISS-3 classified image. One reason for this effect is that NN does not assume data to be normally distributed and DTA uses a non-parametric classification algorithm that involves a recursive partitioning of the feature space, based on a set of rules learned by an analysis of the training set.
Although, hard classification techniques performs well for classifying the image into broader land cover categories having homogeneous surface features, the abundance maps generated from MODIS
data is definitely better when considering mixed pixels. The result is evident from the accuracy assessment performed at the taluk level comparison of the percentage area and the accuracy assessment done at pixel level (showing good agreement with the LISS-3 classified image as discussed above).
Hyperspectral sensing holds the potential to provide a high spectral data obtained about the earth's surface features. The classification of MODIS data (of higher spectral and coarse spatial resolution) shows comparable classification accuracy with LISS-3 MSS data (lower spectral and high spatial resolution) for land cover mapping at a regional level. It is true that the subpixel classification is better than hard classifiers for coarse resolution satellite data, yet to what extent this statement holds good and how far this technique is valid at a global level mapping with high number of endmembers can be a subject of research. Methods can be developed to estimate the accuracy of the fraction images obtained from LMM other than ground survey. Use of high resolution classified maps to extract endmembers including techniques like Iterative Error Analysis (IEA), N-FINDR, etc. for unmixing of MODIS data is also another area of research along with methods on how to optimize the analysis of the high volumes of data acquired. The scope of digital image classification and its application with hyperspectral data is virtually limitless.
Use of neighbourhood operations on classification with the presence of adjacency effects between and within features present on the ground also contribute to heterogeneity and leads to misclassification. The most important research area that has not been studied enough is the Non-Linear Mixture Model (NLMM). The NLMM takes into account the multiple radiances of the ground cover materials, and thus the mixture is no longer linear. The NLMM has a relatively more accurate simulation of physical phenomena and there is not a simple and generic NLMM that can be utilized in various spectral unmixing applications. NLMM can further be investigated and is the subject of ongoing research.
Study of vegetation requires species level identification which could be carried out with much higher spectral and spatial resolution. Future research can be carried out to realise the effect of fusion of high spatial resolution data and true hyperspectral data (having wavelengths in continuous EM spectrum). Use of temporal data can further help in understanding land cover dynamics analysis where the complexities of abrupt policy decisions and lack of prior planning have resulted in environmental problems like haphazard urban sprawl, global warming, climate change etc.