Uni/Multi-Temporal Modified Vegetation Indices
Based on the acquired Sentinel-2A image data for six periods, the average spectral values of seven vegetation types were statistically analyzed (
Figure 3). The spectral differences between the surface features were analyzed according to the spectral feature curves. During February and December, the spectral curve of
Pinus tabuliformis, which is an evergreen plant, exhibited unique vegetation characteristics such as “two valleys and one peak” and “red edge” [
19]. In the Sentinel-2 band, this was represented by the two absorption valleys such as “red valley” and “blue valley”, which were formed by the vegetation near the red and blue bands. Near the green band, a reflection peak was formed, owing to the strong reflection effect on green light, which is also referred to as “green peak”. This feature is due to the influence of chlorophyll, which has a strong absorption effect on blue and red lights and a strong reflection effect on green light. The spectral values of vegetation in the RE2, RE3, and NIR bands sharply increased, forming a nearly straight curve with a large slope, which is commonly known as “red-edge”. This feature depends on the cell structure of the leaves. The other vegetation types are in a dormancy state, with their leaves being withered or with their vegetation spectral features being diminished and their spectral curves tending to be flat, with no evident “peaks” and “valleys”; otherwise, the spectral features are similar to those of bare soil, and they are represented in the Sentinel-2 band as a reflection peak near the short-wave infrared (SWIR). Therefore, there exist significant spectral differences between
Pinus tabuliformis and the other vegetation types. Similarly, as the sowing of major crops did not start in June, there was no vegetation growth in this area. Therefore, the spectral features of its area should be similar to those of bare soil. In contrast, other vegetation had already entered the vigorous growth period, and their spectral curves exhibited evident features of “two valleys and one peak” and “red edge”. Therefore, there exist significant spectral differences between the major crops and other vegetation types. These spectral differences occur because the spectrum of the vegetation growth period is unique and significant, which can be represented by the Sentinel-2 red band and NIR band. These typical vegetation indices were developed for the advantage of increasing the wavelength from visible red light to reflected infrared, which is caused by the selective absorption of red light by chlorophyll for photosynthesis in the vegetation spectrum. Therefore,
Pinus tabuliformis and crops can be easily identified using these typical vegetation indices. In the six time-phases that were selected in this study, scrub grass,
Quercus wutaishanica, pine-oak mixed forests,
Larix principis-rupprechtii, and shaw had relatively similar reflectance values in the red and NIR bands. Therefore, high-precision remote sensing monitoring and identification with typical vegetation indices are difficult. Thus, to improve the classification accuracy of remote sensing vegetation in the LMNR, it is of vital importance to develop a set of modified vegetation indices that can effectively amplify the spectral differences among the scrub grass,
Quercus wutaishanica, pine-oak mixed forests,
Larix principis-rupprechtii, and shaw.
Expanding the inter-class spectral differences and reducing the intra-class differences were the basic principles in developing our modified vegetation indices. Considering the spectral similarities in the five vegetation types that were to be identified, identifying them using a certain vegetation index was difficult, however, the hierarchical classification can effectively improve the classification accuracy [
33]. Therefore, in a sequence from difficult to easy, we developed the indices one by one for specific vegetation types and identified them sequentially. First, through a spectral analysis, we selected the vegetation type with unique spectral features, which was the one with the highest spectral difference among the other vegetation types that were to be identified, which was the object of developing the modified vegetation index. This spectral difference can be represented by dual-band combinations of the same temporal phase or those of different temporal phases, and these dual-band combinations are referred to as feature band combinations. The vegetation may have multiple sets of feature band combinations. Based on standard deviations, multiple sets of feature band combinations were screened, and the band combination with the most stable inter-class difference (i.e., the band combination with the smallest sum of standard deviations) was selected. The calculation formula of the modified vegetation index was determined based on the relationship between the spectral values of the feature dual bands and on the band combinations of typical vegetation (
Figure 4).
In this study, we developed four uni (multi)-temporal modified vegetation indices for pine-oak mixed forests, Quercus wutaishanica, scrub grass, and shaw, and the development process of the four indices is described below.
Regarding, the uni-temporal modified difference pine-oak mixed forest index (UTM-DMI), in February and December, the spectral values of the RE4 band of pine-oak mixed forests are basically the same as those of the SWIR-1 band, and the spectral values of the RE4 band of other vegetation types that were to be identified are significantly lower than those of the SWIR-1 band. This is represented by the reflection peaks near the SWIR-1 and SWIR-2 bands. Therefore, the RE4 and SWIR-1 bands in February and the RE4 and SWIR-1 bands in December are combinations of feature bands that are used for identifying the pine-oak mixed forest images. Based on standard deviations, the two band combinations were further screened, and the band combination with the most stable intra-class variance (i.e., the band combination with the smallest sum of standard deviations) was selected, namely, the RE4 and SWIR-1 bands in February (
Table 1). Based on the principle for constructing the DVI, the UTM-DMI was constructed by calculating the spectral difference between the RE4 band and the SWIR-1 band in February.
where RE4
&Feb denotes the reflectance of the RE4 band in February, and SWIR-1
&Feb denotes the reflectance of the SWIR-1 in February.
Here, the vegetation types that were to be identified include Quercus wutaishanica, scrub grass, Larix principis-rupprechtii, and shaw.
Regarding the multi-temporal modified normalized difference
Quercus wutaishanica Index (MTM-NDQI), in June and October, the spectral values of the NIR and RE4 bands of
Quercus wutaishanica were significantly higher than those of other vegetation types that were to be identified. This is due to the internal structure and density of the
Quercus wutaishanica leaves. Therefore, the NIR band in June and October and the RE4 band in June and October are the combinations of feature bands that were used for identifying the
Quercus wutaishanica images. Based on standard deviations, the two band combinations were further screened, and the band combination with the most stable intra-class variance was selected, namely, the RE4 band in June and October. For different vegetation types, the spectral value of the RE4 band is smaller than 0.54 in June, and it is smaller than 0.34 in October, with their sum being smaller than 1 (
Table 2). The vegetation index variability can be amplified by increasing the numerator value and reducing the denominator value. Based on the principle for constructing the NDVI, the MTM-NDQI was determined using the following equation:
where RE4
&June denotes the reflectance of the RE4 band in June, and RE4
&Oct denotes the reflectance of the RE4 band in October.
Here, the vegetation types that were to be identified include scrub grass, Larix principis-rupprechtii, and shaw.
Regarding the multi-temporal modified difference scrub grass index (MTM-DSI), from June to August, the spectral values of the RE2, RE3, NIR, and RE4 bands of the scrub grass increased, whereas those of other vegetation types that were to be identified decreased. Therefore, the RE2, RE3, NIR, and RE4 bands in June and August constitute four combinations of feature bands for identifying scrub grass images. Based on standard deviations, the four band combinations were further screened, and the band combination with the most stable intra-class variance was selected, namely, the RE2 band in June and August (
Table 3). Based on the principle for constructing the DVI, the MTM-DSI was constructed by calculating the spectral difference between the RE2 band in June and August:
where RE2
&June denotes the reflectance of the RE2 band in June, and RE2
&Aug denotes the reflectance of the RE2 band in August.
Here, the vegetation types that were to be identified include Larix principis-rupprechtii and shaw.
Regarding the multi-temporal modified ratio shaw index (MTM-RSI), the spectral values of the RE2, RE3, NIR, and RE4 bands of shaw are larger than those of
Larix principis-rupprechtii in June, but they are smaller than those of
Larix principis-rupprechtii in October, indicating that from June to October, the spectral values of these bands of shaw changed more significantly than those of
Larix principis-rupprechtii. Therefore, the RE2, RE3, NIR, and RE4 bands in June and October constitute four combinations of feature bands for identifying shaw images. Based on standard deviations, the four band combinations were further screened, and the band combination with the most stable intra-class variance was selected, namely, the RE2 band in June and October (
Table 4). The vegetation index variability can be amplified by increasing the numerator value and reducing the denominator value. Based on the principle for constructing the RVI, the MTM-RSI was constructed by dividing the spectral difference between the RE2 band in June and October and the spectral value of the RE2 band in October:
where RE2
&June denotes the reflectance of the RE2 band in June, and RE2
&Oct denotes the reflectance of the RE2 band in October.
Optimal Feature Set
Figure 5 shows a box plot of typical vegetation indices for seven sample types in the six periods. The overall separability of the NDVI samples is the most significant in February, April, June, and December, and the overall separability of DVI samples is the most significant in August and October. In February, April, and June, the NDVI of crops is lower than that of other vegetation types, while in April, the NDVI of crops tends to overlap less often with that of other vegetation types, making it suitable to identify crop images. In February and December, the NDVI of
Pinus tabuliformis is significantly higher than that of other vegetation types, while in December, the NDVI outliers of
Pinus tabuliformis are less common, making it suitable to identify
Pinus tabuliformis images. In each time phase, the typical vegetation indices of any of the five other vegetation types (including scrub grass,
Quercus wutaishanica, pine-oak mixed forests,
Larix principis-rupprechtii, and shaw) overlap with another vegetation type; the overlapping samples account for more than 25% of the total samples, with low separability being observed between them.
In summary, the NDVI in December and the NDVI in April can easily identify Pinus tabuliformis and crops. Hence, the NDVI in December, the NDVI in April, and the UTM-DMI, MTM-NDQI, MTM-DSI, and MTM-RSI constitute the optimal feature set for remote sensing vegetation classification in the LMNR.