![Self Organizing Maps
𝜔𝑗 = [𝜔𝑗1, …,𝜔𝑗𝑛]𝑥 𝑡 = [𝑥(𝑡)1, …,𝑥(𝑡) 𝑛]](https://image.slidesharecdn.com/2017qualificacao-170905030121/85/Land-Cover-and-Land-use-Classifiction-from-Satellite-Image-Time-Series-Data-using-Self-Organizing-Maps-23-320.jpg)
Land Cover and Land use Classifiction from Satellite Image Time Series Data using Self Organizing Maps
- 1. Land Cover and Land Use Classification From Satellite Image Time Series Data Using Self Organizing Maps Lorena Santos
- 2. Goal This work aims at presenting a literature review on Land Use and Cover Change (LUCC) classification using remote sensing image time series and experiment results using Self Organizing Maps (SOM) method in this context.
- 3. Motivation “The e-sensing project develops new ways to extract information on land use and land cover change from big Earth Observation data sets, using open Science”. Source: Gilberto Camara Remote visualization and method development Big data EO management and analysis How to classify LUCC change using satellite image time series?
- 4. Motivation
- 5. Motivation Test and evaluate new algorithms using neural network for LUCC classification.
- 6. Physical material at the surface of the Earth observed in diferente moments of time Field survey and analysis of remotely sensed imagery are two main methods for obtaining information on land cover (COMBER et.al., 2008) Land cover Tropical forest Water
- 7. Land use Set of activities realized for human on the land cover Land Use are fundamental to planning and implantation of public polices in different scales (COMBER et.al., 2008) Agriculture Logging
- 8. Impacts on tropical ecosystems Increase gases emissions Reduce the planet’s biodiversity Source: Adpated http://www.gly.uga.edu/railsback/CTW2.html Impacts caused by land use and land cover changes
- 9. Challenge To improve the quality of LUCC products, recent research show a growing trend on the use of satellite image time series (Gomez et.al., 2016). Source: www.earthobservation.ie/
- 10. Satellite image time series Phenological information is usually captured by composed indexes such as NDVI and EVI. (FENSHOLT et al., 2015) Vegetation Indices characterize vegetation dynamics across different temporal scales
- 11. Satellite image time series Time series represents a collection of values obtained from sequential measurements over time (ESLING et.al, 2012) In the literature, there are different conceptual approaches for satellite image time series analysis: a) Harmonic analysis; b) Parametric seasonal profiles; c) Pattern matching d) Time series clustering
- 12. Analysis methods for remote sensing time series Clustering Classification Pattern Matching Novelty Detection Source: Camara (2016),Keogh (2006)
- 13. BFAST Source: (VERBESSELT et al., 2010b). Decomposition of time series in: Singular trend models can not be sufficient to identify slower changes. Decomposition allows capture specific vegetation conditions or stages of degradation through time. (FENSHOLT et.al, 2015)
- 14. TIMESAT (a) begin of season; (b) end of season; (c) length of season; (d) base value; (e) time of middle of season; (f) maximum value; (g) amplitude; (h) small integrated value; (i) large integrated value. Source: (KUENZER et al., 2015) Capture metrics that define a plant’s vegetative cycle.
- 15. PATTERN MATCHING Source: (ESLING; AGON, 2012) Finding every subsequence that appears recurrently in a longer time series. These subsequences are named motifs. LUCC time series can be analyzed using pattern matching from motifs.
- 16. PATTERN MATCHING PastureForest Agriculture Source: (MAUS, 2016) Distinct shapes for land patterns A good match needs shape similarity and temporal coherence Forest Forest Agriculture
- 17. Earth Observation Compare two time series and finds their optimal alignment, providing a dissimilarity measure as a result even if they are irregularly sampled or are out of phase in the time axis. Given two time series: Compute a cost matrix ψ, n x m, given by the squared distance between the elements of two time series. ψ𝑖,𝑗 = ( q𝑖, − c𝑗,) 2 Dynamic Time Warping
- 18. Dynamic Time Warping From ψ , we can find the best matching between two time series, getting an optimal path that minimizes the cost warping Resulting alignment
- 19. Earth Observation (Maus, et.al, 2016) introduces a temporal cost, 𝝎𝒊,𝒋, that helps to distinguish between different types o land use and land cover classes. To compute the temporal cost, they propose linear and logistic model 𝑔(𝑡𝑖, 𝑡𝑗) is the elapsed time between the dates in pattern and in time series. Time-Weighted Dynamic Time Warping
- 20. Time-Weighted Dynamic Time Warping Source: (MAUS, 2016) The result of the algorithm is a set of subintervals, each associated with a pattern and with a dissimilarity measure
- 21. Clustering time series How many patterns are required to account for all 17 million time series that are considered to be forest? Time series clustering techniques for big Earth observation data have to be able to map thousands of time series to a few prototypes. Clustering can help to assess whether two time series with the same label in fact match the same pattern.
- 22. Self Organizing Maps SOM methods allow mapping from a high-dimensional space to a low- dimensional space, preserving data topology of data while reducing computational cost Two-dimensional grid representing neurons with their respective models
- 23. Self Organizing Maps 𝜔𝑗 = [𝜔𝑗1, …,𝜔𝑗𝑛]𝑥 𝑡 = [𝑥(𝑡)1, …,𝑥(𝑡) 𝑛]
- 24. Self Organizing Maps For each neuron j, compute Euclidean Distance between vector x and weight vector 𝜔𝑗 Best Matching Unit Competitive Process
- 25. Self Organizing Maps The topological neighborhood depends on lateral distance, 𝑺 𝒃𝒋, between the winner neuron b and neuron j. σ determines how the size of neighborhood decrease in time Cooperative Process
- 26. Self Organizing Maps Winning neuron must be updated at each iteration Adaptive Process The learning rate 𝜶(𝒕) decreases gradually with time t An iteration ends when all vectores of input layer are trained.
- 27. Hierarchical Clustering Earth Observation Agglomerative Start with leaves and aggregate Works by grouping data objects into a tree of clusters P Q R S Divisive Start root and subdivide
- 28. Hierarchical Clustering Earth Observation Agglomerative Start with leaves and aggregate Works by grouping data objects into a tree of clusters P Q R S Divisive Start root and subdivide
- 29. Hierarchical Clustering Earth Observation Agglomerative Start with leaves and aggregate Works by grouping data objects into a tree of clusters P Q R S Divisive Start root and subdivide
- 30. Hierarchical Clustering Earth Observation Agglomerative Start with leaves and aggregate Works by grouping data objects into a tree of clusters P Q R S Divisive Start root and subdivide
- 31. Hierarchical Clustering Earth Observation Agglomerative Start with leaves and aggregate Works by grouping data objects into a tree of clusters P Q R S Divisive Start root and subdivide P Q R S
- 32. Hierarchical Clustering Earth Observation 3 1 5 2 4 Single linkage A B Earth Observation 3 1 5 2 4 Complete linkage A B
- 33. Hierarchical Clustering Earth Observation 3 1 5 2 4 Average linkage A B 𝐷𝑖𝑠𝑡(𝐴, 𝐵) = 𝑑1,4+𝑑1,5+ 𝑑2,4+ 𝑑2,5+ 𝑑3,4+𝑑3,5 6 Ward’s linkage Merges two clusters that result in the smallest increase in the value of the sum-of-square variance. The sum-of-square variance is computed for each cluster, and the on with the smallest value is selected.
- 34. Hierarchical Clustering 1 3 2 5 4 6 0 0.05 0.1 0.15 0.2 Dendrogram helps to evaluate the height where the largest change in dissimilarity occurs. Dendrogram: a tree-like diagram that records the sequences of merges or splits
- 35. LUCC classification using remote sensing time series LUCC classification for image time series consists of two steps: a) Prototype definition : assign each member of the training set to one cluster or exemplars that representes a separable set of input data. b) Time series classification: after a set of clusters, prototypes or examplars are defined, they are used to classify all time series in the input data set.
- 36. LUCC classification using remote sensing time series Author Method Result Morse-McNabb et.al Decision tree 75% accuracy Liu et.al Decison tree 80% accuracy Arvor et.al Maximum Likelihood 74% accuracy Bendini et.al Radom forest 95% cross-validation Xue et.al Ensemble learning + SVM 95% accuracy
- 37. LUCC classification from satellite image time series using SOM Employed SOM neural network technique to classify land cover types in easter China during growing period of plants. Adress three aspects: 1. To evaluate the ability of SOM to classify land cover by using high dimensional time-series data. 2. To compare the results by SOM method with de maximum likedhood classifier. Conclusions: This study comes to a major conclusion that SOM neural network is a more effective method for land cover classification compared with traditional MLC method by using high-dimensional time-series MODIS data.
- 38. Source: (Bagan et.al, 2005)
- 39. Source: (Bagan et.al, 2005) LUCC classification from satellite image time series using SOM
- 40. Source: (Bagan et.al, 2005) LUCC classification from satellite image time series using SOM
- 41. Case study
- 42. Case study
- 43. Case study Class Count Forest 138 Cotton-Fallow 68 Soybean-Cotton 79 Soybean-Maize 134 Soybean-Millet 184 Total 603
- 44. Kohonen Maps Parameters SOM: Input vectors: 603 time series Output layer: 7 x 7 Iterations: 1000 Learning rate: α(0)= 0.1 and α(t)= 0.02
- 46. Forest
- 47. Cotton fallow
- 48. Soybean millet
- 49. Soybean Maize
- 50. Soybean Cotton
- 51. Final Remarks The use of remote sensing time seriers have provide better characterization of land cover dynamics. We are starting to explore new methods for LUCC classification using time series. SOM was the first experiment and conclude that SOM method is suitable for the following taks : a) To assess land use and land cover samples. b) To create signatures of land use and land cover classes. c) For LUCC classification.