Cartografía automática de área quemada a nivel local-regional mediante algoritmos de contexto espacial

  1. BASTARRICA IZAGUIRRE, AITOR
Supervised by:
  1. Emilio Chuvieco Salinero Director
  2. M. Pilar Martín Isabel Director

Defence university: Universidad de Alcalá

Fecha de defensa: 21 September 2009

Committee:
  1. F. López Chair
  2. Inmaculada Aguado Suárez Secretary
  3. Olga Viedma Sillero Committee member
  4. María Cristina Vega García Committee member
  5. Jesús San Miguel Ayanz Committee member

Type: Thesis

Teseo: 282297 DIALNET lock_openTESEO editor

Abstract

The core of the present doctoral thesis is the development of a methodology to map burned land-areas in the Mediterranean region automatically. It is applicable upon common images used for mapping burned areas, at various scales, like TM/ETM+ (medium space resolution) and MODIS (low space resolution). The methodology proposed is based upon the so-called multi-phase algorithms, where mapping proceeds through two stages. Two-stage algorithms allow overcoming problems related to single-stage procedures: the variability shown by the burned signal depending on the fire that caused them (i.e. fire intensity, propagation speed, etc.), the type and state of the vegetation previous to the occurrence of the fire (i.e. biomass, horizontal and vertical continuity), and the time elapsed between the fire and the acquisition of the image. In the first stage the identification of the areas presenting the highest probability of being burned (called seeds) is carried out while trying to minimize commission errors. In the second stage, and commencing from the seeds identified in the previous stage, less strict contextual criteria allow defining the perimeters of the burnt regions while trying to reduce the omission errors. The first stage of the proposed algorithm has proved to be much more important than the second one due to the fact that non-identified real fires cannot be mapped. Also, the inclusion of false fires can lead to defining erroneous perimeters depending on the similarity criteria used in the second stage. The rules applied in the first stage were the result of classification trees that highlighted the relevance of having two bands in the SWIR to obtain an automatic precise identification of burned areas. The rules proposed to identify the burned areas on TM/ETM+images rested upon two post-fire indices: Burned Area Index (BAIM) and Mid Infrared Burned Index (MIRBI). Inclusion of data containing information previous to the fire has not been necessary. When using MODIS data, however, the temporal component has shown to be critical and the chosen rules were mainly based on the changes in time caused by the fire in the variables BAIM, Normalized Burned Ratio (NBR) and MIRBI, where temporal windows of 3, 5 and 7 week, respectively, have been applied. The variable used by the region-growing algorithm was derived from the probability values returned by a multi-temporal logistic regression model. This model has shown to be highly capable to statistically distinguish ‘burned’ and ‘not-burned’ category on TM/ETM+ and MODIS images. It is this capability of clear distinction the reason why different tested contextual algorithms provide similar results. Finally, the algorithm called Fijo+bordes has been chosen as the most appropriate: it presents the best agreement with the reference data and it has demonstrated a highly computational efficiency that eases its implementation within a burnedscar cartography system. The chosen algorithm takes into account the local variability between burned and not-burned areas previously defined by the Sobel edge detector. In addition to this criteria the algorithm uses a fixed probability threshold value to stop the region growing process started from the seeds set in the first stage. Edge detectors are helpful tools due to their spatial heterogeneity results on a bigger number of edges: they avoid growing of areas burnt on previous years or areas that present a high probability value even though they have not been burned. Nonetheless, edge detectors omit fragmented burnt areas: the borders make the growing of the seeds to reach the perimeter of the area more difficult. This is the reason why the algorithm is more efficient when the burnt-areas are more compact and spectrally more homogeneous. The results of the developed automatic algorithm have shown, for the 18 study areas assessed with TM/ETM+ images a very good agreement with the reference data generated with those same images. For the case of MODIS data the algorithm has been assessed in three Mediterranean regions and the agreement with reference data --released by the European Forest Fire Information System (EFFIS) and other official institutions (case of Portugal)-- is very good. The proposed algorithm has proved to be more efficient than the version 5 MODIS Burned Area Monthly L3 Global 500m (MCD45A1) product.