Unsupervised validationThe performa

Unsupervised validation
The performance of coordinates V and W respectively as a classifier
of vegetated surfaces and as a proxy of water content may be assessedby means of an unsupervised technique that allows distinguishing
among classes such that the between-group variance of the specified
number of classes is maximized. A commonly used unsupervised
technique is the one know as K-means or C-means (MacQueen, 1967)
which essentially consists in prescribing an initial cluster centre forgreener patches in both scenes, the soil-stressed/sparse vegetation
may be identified as the pinkish and purple areas; finally, the burned
surfaces are readily identifiable as the very dark or black pixels of the
ETM+images. As expected, in the η/ξ space, the grey pixels (i.e. those
belonging to the “other” types cluster do not stand close to coordinate
curve V=1, as opposed to the “vegetated” type cluster, whose pixels
lie along that coordinate curve.
A summary of results of K-means for all 16 scenes is presented in
Table 4 and the obtained overall consistency is worth being noted. In
all 16 scenes analyzed the V cluster with centroid around 0.97 to 0.99
is associated with vegetated surfaces containing. As expected, the
other V cluster is less stable, since it considerably depends upon the
“other” types of surface (e.g. clouds or water bodies) that is present in
the image. The centres of the W clusters also depend on the types of
landcover in each scene and, for this reason; results have to be
compared against the high resolution image taken on the same day.
Accordingly, scenes 1, 2, 3, 4, 5, 6, 7, 10 and 14, that contain burned
areas always have the cluster with centre of lowest value (close to
0.1). On the other hand, scenes mostly covered by vegetation, usually
have the cluster with centre of highest value (about 0.23). Finally, soil
and sparsely vegetated areas are
each of the sought-after clusters and then assigning each pixel of the
set to the class nearest to the pixel. A new cluster centre is calculated
and pixels are reassigned accordingly. The procedure is repeated until
no significant changes in pixel assignments occur from a given
iteration to the next.
Taking into account the different characteristics of V and W that
make of them respectively a good classifier and a good quantifier, the
K-means algorithm was successively applied to coordinates V and W
of several MODIS images; i) two cluster centers were estimated from
the V sample (respectively associated to vegetated surfaces and to
other types) and ii) four clusters were then derived from the W
sample restricted to those pixels belonging to the cluster associated to
vegetated surfaces (i.e. the one with centre of higher V). Results
obtained from the unsupervised classification of each image were
finally compared against Landsat ETM+high resolution image taken
on the same day (see Table 1).
Figs. 16 and 17 present the results obtained after applying Kmeans to scenes 3 and 4, respectively. Regarding to the η/ξ space (left
panel), gray points correspond to the first of the two clusters obtained
by applying K-means to V whereas colored points represent the
second cluster. This second cluster was then used as input to a second
K-means procedure which was applied to coordinate W. Thus each
colored cluster denotes the clusters derived from the K-means from
W, as suggested by the drawn contour lines indicating the limits
between these clusters. It is worth noting that colors in the left and
central panels correspond to the same clusters. Taking for reference
the RGB (543) of the high resolution images (Figs. 16 and 17, right
panels), it may be visually confirmed that, when applied to the V
samples, the K-means algorithm is able to discriminate between
pixels associated to vegetated surfaces (green vegetation, stressed
vegetation, and burned surfaces), on the one hand and to the other
non-vegetated types (e.g. water bodies and clouds), on the other. The
two clusters, whose centres respectively present a high and a low
value of V, will be hereafter referred to as vegetated and “other” types.
“Other” pixels correspond therefore to the gray points in the left
panels of Figs. 16 and 17, whereas the remaining colors identify the
pixels belonging to the vegetated type. When K-means is further
applied to vegetated pixels, the obtained four clusters in W appear to
be related respectively to one class of green vegetation (represented
in green), two classes of soil or dry vegetation or sparsely vegetated
areas (represented in dark green and dark brown) and one class of
burned surfaces (represented in black). A close agreement may be
visually identified between the spatial patterns of the above-referred
five classes (central panels) and the spatial distribution of RGB (543)
pixels (right panels). For instance, the “other” types cluster corresponds to clouds in case of scene number 3 (Fig. 16) and to water in
case of scene 4 (Fig. 17); the green vegetation class corresponds to the

Hasil (Bahasa Indonesia) 1:

Validasi tanpa pengawasanKinerja koordinat V dan W masing-masing sebagai classifierpermukaan tumbuhan dan sebagai proxy kadar air mungkin assessedby sarana teknik tanpa pengawasan yang memungkinkan membedakangolongan sedemikian rupa sehingga varians antara kelompok yang ditetapkanJumlah kelas yang maksimal. Umum digunakan tanpa pengawasanteknik adalah salah satu tahu sebagai K-cara atau C-sarana (MacQueen, 1967)yang pada dasarnya terdiri dalam resep pusat awal gugus forgreener patch dalam adegan kedua, tanah-menekankan/vegetasimungkin diidentifikasi sebagai daerah kemerahan dan ungu; Akhirnya, dibakarpermukaan mudah diidentifikasi sebagai sangat gelap atau hitam pixel dariETM + gambar. Seperti yang diharapkan, di ruang η ξ, abu-abu piksel (yaitu merekamilik gugus jenis "lain" tidak berdiri dekat untuk mengkoordinasikankurva V = 1, sebagai lawan jenis "tumbuhan" cluster, piksel yangTerletak di sepanjang kurva koordinat.Ringkasan hasil K-berarti untuk semua adegan 16 disajikan dalamTabel 4 dan konsistensi keseluruhan diperoleh bernilai karena tercatat. DalamSemua 16 adegan dianalisis cluster V dengan centroid di sekitar 0.97 untuk 0,99ini dikaitkan dengan permukaan bervegetasi yang mengandung. Seperti yang diharapkan,cluster V lain kurang stabil, karena itu sangat tergantung pada"lain" jenis permukaan (misalnya awan atau badan air) yang hadir dalamgambar. Pusat kluster W juga tergantung pada jenispenutupan lahan di setiap adegan, dan untuk alasan ini; hasil harusdibandingkan dengan gambar resolusi tinggi yang diambil pada hari yang sama.Dengan demikian, adegan 1, 2, 3, 4, 5, 6, 7, 10 dan 14, yang mengandung dibakardaerah selalu memiliki cluster dengan pusat nilai terendah (dekat dengan0.1). di sisi lain, adegan sebagian besar tertutup oleh vegetasi, biasanyamemiliki gugus dengan pusat nilai tertinggi (sekitar 0,23). Akhirnya, tanahdan jarang bervegetasi daerahsetiap cluster dicari dan kemudian menetapkan setiap pixel darimenetapkan kelas terdekat pixel. Pusat kluster baru dihitungdan piksel dipindahkan sesuai. Prosedur ini diulang sampaiada perubahan yang signifikan dalam pixel tugas terjadi dari diberikaniterasi berikutnya.Dengan memperhatikan berbagai karakteristik V dan W yangmembuat mereka masing-masing classifier baik dan quantifier baik,K-berarti algoritma berturut-turut diterapkan ke koordinat V dan Wbeberapa MODIS gambar; i) dua gugus pusat diperkirakan darisampel V (masing-masing terkait ke permukaan tumbuhan dan untukjenis lain) dan ii) empat kluster kemudian diperolehi dari Wsampel terbatas piksel tersebut milik klaster yang terkaittumbuhan permukaan (yaitu satu dengan pusat v lebih tinggi). HasilDiperoleh dari klasifikasi tanpa pengawasan setiap gambar yangakhirnya dibandingkan terhadap Landsat ETM + gambar resolusi tinggi diambilon the same day (see Table 1).Figs. 16 and 17 present the results obtained after applying Kmeans to scenes 3 and 4, respectively. Regarding to the η/ξ space (leftpanel), gray points correspond to the first of the two clusters obtainedby applying K-means to V whereas colored points represent thesecond cluster. This second cluster was then used as input to a secondK-means procedure which was applied to coordinate W. Thus eachcolored cluster denotes the clusters derived from the K-means fromW, as suggested by the drawn contour lines indicating the limitsbetween these clusters. It is worth noting that colors in the left andcentral panels correspond to the same clusters. Taking for referencethe RGB (543) of the high resolution images (Figs. 16 and 17, rightpanels), it may be visually confirmed that, when applied to the Vsamples, the K-means algorithm is able to discriminate betweenpixels associated to vegetated surfaces (green vegetation, stressedvegetation, and burned surfaces), on the one hand and to the othernon-vegetated types (e.g. water bodies and clouds), on the other. Thetwo clusters, whose centres respectively present a high and a lowvalue of V, will be hereafter referred to as vegetated and “other” types.“Other” pixels correspond therefore to the gray points in the leftpanels of Figs. 16 and 17, whereas the remaining colors identify thepixels belonging to the vegetated type. When K-means is furtherapplied to vegetated pixels, the obtained four clusters in W appear tobe related respectively to one class of green vegetation (representedin green), two classes of soil or dry vegetation or sparsely vegetatedareas (represented in dark green and dark brown) and one class ofburned surfaces (represented in black). A close agreement may bevisually identified between the spatial patterns of the above-referredfive classes (central panels) and the spatial distribution of RGB (543)pixels (right panels). For instance, the “other” types cluster corresponds to clouds in case of scene number 3 (Fig. 16) and to water incase of scene 4 (Fig. 17); the green vegetation class corresponds to the