ObjectiveThe classification of forest fuels based on high resolution remote sensing images is very important for the modern management of forest fire, but the study for classification by multi-temporal high resolution remote sensing images is scanty at home and abroad. This study explored the classification method of high-resolution images, the differences in classification results of multi-temporal forest fuels, and researched the relationship with altitude and slope change.
MethodAccording to the vegetation status and previous research results in the Jiufeng Forest Farm of Beijing, the fuels were classified by plant community, forest types and combustion characteristics. Then we studied and compared the spectral characteristic curves of different forest fuel types, and established the connection between remote sensing images and forest fuels. The remote sensing images of May, August and October of GF-1 were used as the original data. The classification of forest fuels was carried out by the support vector machine (SVM) algorithm, random forest (RF) and the decision tree method based on CART of EnMAP-box in the Jiufeng Forest Farm, and the classification results were as follows: coniferous forest, broadleaved forest, coniferous and broadleaved mixed forest, shrub forest and non-forest land. After describing their characteristics separately, the optimal classification method was applied into multi-temporal remote sensing images, and the change detection algorithm was used to determine the changes among the types of forest fuels during non-fireproof period (May to November). At the same time, we divided the digital elevation model (DEM) into four categories (1 (< 250 m), 2 (250-500 m), 3 (500-750 m) and 4 (>750 m)). Similarly, we divided slope into three types: gentle slope (< 15°), slope (15°-35°), steep slope (>35°), and used the Jenks method to calculate the percentage of land area change for each category of elevation and slope, respectively. Then we studied the changes in the classification results of forest fuels with changes of altitude and slope.
ResultThe results showed that the spectral characteristics of the five forest fuel categories were well differentiated. The SVM classification was the most accurate. The penalty parameter (C) was 1 000 and the kernel parameter (g) was 10, which made the SVM classification model optimal. The overall classification accuracy was 91.88%, the kappa coefficient was 0.89. And the accuracy was improved relative to RF and CART. The classification accuracy was 2.72% and 9.36% higher than RF and CART, respectively. The types of forest fuels during non-fireproof period (May to November) had certain change regularity, and there were no significant changes in coniferous and mixed forest which belong to moderate stable types, keeping 93.74% and 94.87%, respectively. In contrast, broadleaved and shrub forest changed greatly by 14.64% and 13.36%, respectively; with the increase of altitude and the change of slope, the land area of forest fuels had also changed. The area with altitude above 500-750 m and slope of 16°-35° had the largest change, reaching more than 20%.
ConclusionIn the classification of forest fuels with multi-temporal high-resolution remote sensing images, the SVM classification method can classify fuels better, and with the change of time, altitude and slope, the change of forest fuel area has certain regularity. From May to October, broadleaved forests and shrubs vary most at altitudes of 500-750 m and slopes of 16°-35°.
DownLoad: