Python Machine Learning Conference & GeoPython 2020

»Analysis of burnt scar using optical and radar satellite data«
2020-09-21, 13:45–14:15, Room 2

To analyze the use of satellite SAR data and its comparison to optical imagery for identification and classification of burnt and unburnt patches after a forest fire.

This study compares the use of Sentinel 1 (S1) SAR sensor alongside with Sentinel 2 (S2) optical sensor in detection and mapping of burnt and unburnt scars occurring after a bushfire in Victoria, Australia, and Spain. The bushfires had recently occurred in the period of 2017-2018. The C-band dual polarized S1 data have been investigated to assess the backscatter intensity together with polarimetric decomposition component to determine forest burn severity over the two sites. The backscatter coefficient was also used in deriving texture measures from local statistics, using grey level co-occurrence matrix (GLCM). This was because of its sensitivity in the identification of textural variation of burnt and unburnt scars. While for S2 the difference normalized burnt ratio (dNBR) was utilized to determine the magnitude of burnt severity levels present in both areas. Its analysis was explored using a contextual classifier Support Vector Machine and Markov Random Field classifier (SVM-MRF). This is because of its integration of spectral information and spatial context through the optimal smoothing parameter without degrading image quality. The training and test set datasets consisting of burned and unburned pixels were created from S2 scenes used as reference data. The experimental results showed that a strong correlation exists in both spectral sensitivity and polarimetric sensitivity of the two defined classes after classification. The performance of the algorithm was evaluated using the kappa coefficient and f-score measurement. All fire zones yielded an accuracy of (0.80) except for S1 data in Spain. Also, the performance in users and producers accuracy provided the highest accuracies in both S1 and S2. The entropy alpha decomposition helped to classify the target based on their physical properties as presented by the H-α plane. The entropy and alpha values decreased and formed a pattern after the fire. The sensitivity analysis to the GLCM features showed that homogeneity, contrast and entropy were the key statistical features that showed clear separation of burnt and unburnt scars using backscatter intensity. This was after the key parameters such as number of quantization levels, window size, pixel pair sampling distance which was one and the orientation were optimized. The use of S1 in discrimination of burnt and unburnt scars was highly dependent on local incidence angle, acquisition geometry and environmental conditions. In hilly areas, the low incidence angles showed high discrimination of burnt from unburnt areas compared to high incidence angles. Also, topography was of high influence as areas facing slopes in hilly areas showed high discrimination of unburnt areas from burnt compared to areas facing backslopes. The Spain dataset did not foreshow any changes in vegetation structure after the fire as compared to Australia using S1. This led to the conclusion that also the intensity of the fire and its effect to vegetation structure is of great influence to the sensitivity of SAR sensor in the analysis of changes in forest structure after a bushfire. Also optical data in such cases can be used as a substitute as it showed strong spectral sensitivity to changes in Spain fire irrespective of the intensity of the fire. Nevertheless, results in both areas verify the use of satellite SAR sensor and optical in forestry application and their sensitivity highly depends on vegetation structure, geographical nature of the area of study and fire intensity.