Download as PDF, PPTX
![Get the Target Image
band_a = rasterio.open(image_file_path.format(scene_id, 1))
band_g = rasterio.open(image_file_path.format(scene_id, 3))
x_min, x_max, y_min, y_max = (
354479.71320833557, 395642.8506635118, 2764434.5394466477, 2702420.0095054144
)
gt = band_a.get_transform()
x_min, x_max = floor((x_min - gt[0]) / gt[1]), ceil((x_max - gt[0]) / gt[1])
y_min, y_max = floor((y_min - gt[3]) / gt[5]), ceil((y_max - gt[3]) / gt[5])
band_a, band_g = (
gpuarray.to_gpu(
band_a.read(1, window=((y_min, y_max), (x_min, x_max))).astype(float)
),
gpuarray.to_gpu(
band_g.read(1, window=((y_min, y_max), (x_min, x_max))).astype(float)
)
) GPU](https://image.slidesharecdn.com/gpu-170222115631/75/Landset-8-11-2048.jpg)
![Correction
correct = ElementwiseKernel(
"double *band_a, double *band_g, double *result",
"result[i] = band_a[i] + ((1 / 4.0) * (band_a[i] - band_g[i]) / band_g[i])",
"correct"
)
correct(band_a, band_g, result)
GPU
result = band1 + (1 / 4.0) * (band1 - band3) / band1
CPU](https://image.slidesharecdn.com/gpu-170222115631/75/Landset-8-12-2048.jpg)
![Bubble Sort
cmp_swap = ElementwiseKernel(
"double *x, double *y, char *xx, char *yy",
'''
if(x[i] > y[i]) {
double tmp_d;
char tmp_u;
tmp_d = x[i]; x[i] = y[i]; y[i] = tmp_d;
tmp_u = xx[i]; xx[i] = yy[i]; yy[i] = tmp_u;
}
''',
“cmp_swap")
for j in range(len(images)-1, 0, -1):
for i in range(0, j):
cmp_swap(images[i], images[j], images_index[i], images_index[j])
GPUCPU times: user 1.42 s, sys: 460 ms, total: 1.88 s Wall time: 1.99 s](https://image.slidesharecdn.com/gpu-170222115631/75/Landset-8-13-2048.jpg)
![Bubble Sort
def cmp_swap(x, y, xx, yy):
shape_y, shape_x = x.shape
for i in range(shape_y):
for j in range(shape_x):
if x[i, j] > y[i, j]:
x[i, j] , y[i, j] = y[i, j], x[i, j]
for j in range(len(images)-1, 0, -1):
for i in range(0, j):
cmp_swap(images[i], images[j], images_index[i], images_index[j])
CPUSpend too much time. Details are as “Evaluation”](https://image.slidesharecdn.com/gpu-170222115631/75/Landset-8-14-2048.jpg)
More Related Content
![【JAWS-UG Shimane vol.5 】[ハンズオン]サーバーレスで作るチャットBot](https://cdn.slidesharecdn.com/ss_thumbnails/20160604awshandson-160607034637-thumbnail.jpg?width=640&height=640&fit=bounds)
Landset 8 的雲層去除技巧實作
- 1. CLOUD DETECTION ANDREMOVAL TECHNIQUES FOR LANDSAT 8 IMAGERY WITH GPU ACCELERATION 2016衛星影像處理理 / B10209047 / 倪鈵斯
- 2.
- 3. Abstract 此報告採⽤用之演算法主要是參參考「Cloud Detection andRemoval Techniques for Landsat 8 Imagery」（詳⾒見見「Previous Work」）中， 所介紹之雲層去除技巧，並使⽤用Nvidia CUDA平⾏行行運算框架實作之。 在報告中，學⽣生實作了了兩兩個版本： CPU計算以及GPU計算，⽤用以比較 兩兩者的效能差異異。 VS
- 4.
- 5. Previous Work 在前半段⽂文獻提到ACCA演算法。然⽽而，⽂文獻認為Landset-8的Quality Assessment Bands尚未善⽤用Landset-8新增的頻帶，此外，也無法偵測雲 層所造成的陰影。 在此，它參參考了了⼀一個更更簡單的哲學：針對⼤大量量的影像做排序，取出RGB數 值較暗的⼦子集合，再加以平均。
- 6. Previous Work 有別於直接使⽤用RGB數值，⽂文獻採⽤用Coastal AerosolBand作為排序的 依據，因為此頻帶對雲層更更加敏感。 Band Number Name µm Resolution 1 Coastal / Aerosol 0.433–0.453 30 m 2 Visible Blue 0.450–0.515 30 m 3 Visible Green 0.525–0.600 30 m 4 Visible Red 0.630–0.680 30 m 5 Near Infrared 0.845–0.885 30 m … … … …
- 7.
- 8.
- 9.
- 10.
- 11. Get the TargetImage band_a = rasterio.open(image_file_path.format(scene_id, 1)) band_g = rasterio.open(image_file_path.format(scene_id, 3)) x_min, x_max, y_min, y_max = ( 354479.71320833557, 395642.8506635118, 2764434.5394466477, 2702420.0095054144 ) gt = band_a.get_transform() x_min, x_max = floor((x_min - gt[0]) / gt[1]), ceil((x_max - gt[0]) / gt[1]) y_min, y_max = floor((y_min - gt[3]) / gt[5]), ceil((y_max - gt[3]) / gt[5]) band_a, band_g = ( gpuarray.to_gpu( band_a.read(1, window=((y_min, y_max), (x_min, x_max))).astype(float) ), gpuarray.to_gpu( band_g.read(1, window=((y_min, y_max), (x_min, x_max))).astype(float) ) ) GPU
- 12. Correction correct = ElementwiseKernel( "double*band_a, double *band_g, double *result", "result[i] = band_a[i] + ((1 / 4.0) * (band_a[i] - band_g[i]) / band_g[i])", "correct" ) correct(band_a, band_g, result) GPU result = band1 + (1 / 4.0) * (band1 - band3) / band1 CPU
- 13. Bubble Sort cmp_swap =ElementwiseKernel( "double *x, double *y, char *xx, char *yy", ''' if(x[i] > y[i]) { double tmp_d; char tmp_u; tmp_d = x[i]; x[i] = y[i]; y[i] = tmp_d; tmp_u = xx[i]; xx[i] = yy[i]; yy[i] = tmp_u; } ''', “cmp_swap") for j in range(len(images)-1, 0, -1): for i in range(0, j): cmp_swap(images[i], images[j], images_index[i], images_index[j]) GPUCPU times: user 1.42 s, sys: 460 ms, total: 1.88 s Wall time: 1.99 s
- 14. Bubble Sort def cmp_swap(x,y, xx, yy): shape_y, shape_x = x.shape for i in range(shape_y): for j in range(shape_x): if x[i, j] > y[i, j]: x[i, j] , y[i, j] = y[i, j], x[i, j] for j in range(len(images)-1, 0, -1): for i in range(0, j): cmp_swap(images[i], images[j], images_index[i], images_index[j]) CPUSpend too much time. Details are as “Evaluation”
- 15.
- 16. Evaluation 程式主邏輯時間複雜度⼤大約為 ，m為混合影像數量量、n為影像⾼高度、k為影像長 度。根據此規則，完整排序相同⼤大⼩小的影像，GPU僅需花費1.99秒，⽽而CPU卻需要花5000秒。 秒 0 1250 2500 3750 5000 5 1015 20 25 30 35 40 45 50 55 60 65 70 75 80 85 CPU GPU
- 17. END THANK YOU FORYOUR ATTENTION