lee滤波对于sar影像去噪非常有用

关注

原创最新推荐文章于 2026-10-25 13:01:31 发布 375 阅读 AI 写同款· GEO 优化 ›

Lee滤波和Refine-Lee滤波 Lee滤波是选择一定长度的窗口作为局部区域，利用图像局部统计特性进行SAR图像斑点滤波的典型方法。其是基于完全发育的斑点噪声模型。（图像上斑点处于均匀区域，且斑点噪声与图像信号不相关）根据试验研究表明完全发育的相干噪声是一种乘性噪声，即有(k,l)是图像像素的坐标，I(k,l)是实际得到的图强强度（含有噪声），x(k,l)是一个平稳随机过程，描述了原始信号，v(k,l)即为相干斑噪声，它是一个均值为1，方差为的平稳白噪声。乘性相干斑模型如下：噪声抑制的两个关键环节：一是建立真实后向散射系数的估计机制，二是制阅读详情

lee滤波对于sar影像去噪非常有用。且lee滤波也很容易理解，且看以下python代码编写的lee滤波函数。

import numpy as np
from scipy.ndimage import uniform_filter

def lee_filter(img, size, global_variance):
    """
    Apply Lee filter for speckle noise reduction using a global variance.
    """
    img_mean = uniform_filter(img, (size, size))
    img_sqr_mean = uniform_filter(img ** 2, (size, size))
    img_variance = img_sqr_mean - img_mean ** 2

    # Avoid division by zero or invalid operations
    denominator = img_variance + global_variance
    denominator[denominator == 0] = 1e-9  # Prevent division by zero

    img_weights = img_variance / denominator
    img_output = img_mean + img_weights * (img - img_mean)
    return img_output

但是在实际处理时，我们一般还要做一些处理，比如规定好输入、输出。一般来说，遥感影像是巨大的，所以我们也要考虑到硬件的性能。

以下是我所使用的改进的代码，仅供参考。

import numpy as np
from scipy.ndimage import uniform_filter
from osgeo import gdal
import sys

def translate(band, number=2):
    """
    Convert the value of a band to the range 0-255 for visualization
    2% linear conversion
    """
    if np.all(np.isnan(band)) or not np.any(band):
        return np.zeros_like(band, dtype=np.uint8)

    max_val = np.nanmax(band)
    min_val = np.nanmin(band)

    if max_val == 0:
        return np.zeros_like(band, dtype=np.uint8)

    band = band.astype(np.float64)
    band[band == -9999] = np.nan
    band[band == 0] = np.nan

    band_data = band / max_val * 255
    band_data[np.isnan(band_data)] = 0

    d2 = np.percentile(band_data, number)
    u98 = np.percentile(band_data, 100 - number)

    if u98 - d2 == 0:
        return np.zeros_like(band, dtype=np.uint8)

    maxout = 255
    minout = 0
    data_8bit_new = minout + ((band_data - d2) / (u98 - d2)) * (maxout - minout)
    data_8bit_new[data_8bit_new < minout] = minout
    data_8bit_new[data_8bit_new > maxout] = maxout
    return data_8bit_new

def read_img(file):
    """
    Read a Sentinel-1 GeoTIFF image and return its data, geotransform, projection, and GCPs.
    """
    dataset = gdal.Open(file)
    if dataset is None:
        raise FileNotFoundError(f"Could not open {file}")

    img_width = dataset.RasterXSize
    img_height = dataset.RasterYSize
    img_data = dataset.GetRasterBand(1).ReadAsArray(0, 0, img_width, img_height)
    img_data = img_data.astype(float)
    img_data[img_data == -9999] = np.nan
    img_data[img_data == 0] = np.nan

    geotransform = dataset.GetGeoTransform()
    if geotransform is None:
        raise ValueError("Input image has no geotransform information")

    projection = dataset.GetProjection()


    # Retrieve GCPs and GCP projection
    gcps = dataset.GetGCPs()
    gcp_projection = dataset.GetGCPProjection() if gcps else None

    return img_data, dataset, geotransform, projection, gcps, gcp_projection

def lee_filter(img, size, global_variance):
    """
    Apply Lee filter for speckle noise reduction using a global variance.
    """
    img_mean = uniform_filter(img, (size, size))
    img_sqr_mean = uniform_filter(img ** 2, (size, size))
    img_variance = img_sqr_mean - img_mean ** 2

    denominator = img_variance + global_variance
    denominator[denominator == 0] = 1e-9

    img_weights = img_variance / denominator
    img_output = img_mean + img_weights * (img - img_mean)
    return img_output

def process_block(full_img_data, block_y, block_x, block_height, block_width, window_size, global_variance):
    """
    Process a single block with padding to handle edge effects correctly.
    Expects full_img_data to be pre-filled with 0 instead of NaN.
    """
    full_height, full_width = full_img_data.shape
    pad_size = window_size // 2

    y_start_padded = max(0, block_y - pad_size)
    y_end_padded = min(full_height, block_y + block_height + pad_size)
    x_start_padded = max(0, block_x - pad_size)
    x_end_padded = min(full_width, block_x + block_width + pad_size)

    padded_block = full_img_data[y_start_padded:y_end_padded, x_start_padded:x_end_padded]
    filtered_padded_block = lee_filter(padded_block, window_size, global_variance)

    y_start_crop = block_y - y_start_padded
    y_end_crop = y_start_crop + block_height
    x_start_crop = block_x - x_start_padded
    x_end_crop = x_start_crop + block_width

    return filtered_padded_block[y_start_crop:y_end_crop, x_start_crop:x_end_crop]

def block_processing(img_data, output_path, dataset, geotransform, projection, gcps, gcp_projection, block_size=512, window_size=7):
    """
    Process the image in blocks, replacing NaNs with 0 for filtering and restoring them after.
    Preserves Geo Points (geotransform), Coordinate System (projection), and GCPs in output.
    """
    full_height, full_width = img_data.shape
    driver = gdal.GetDriverByName('GTiff')
    if driver is None:
        raise ValueError("GDAL GTiff driver not available")

    out_dataset = driver.Create(output_path, full_width, full_height, 1, gdal.GDT_Float32)
    if out_dataset is None:
        raise RuntimeError(f"Failed to create output file: {output_path}")

    # Set geospatial metadata
    out_dataset.SetGeoTransform(geotransform)
    out_dataset.SetProjection(projection)
    if gcps and gcp_projection:
        print("Applying GCPs to output image...")
        out_dataset.SetGCPs(gcps, gcp_projection)

    out_band = out_dataset.GetRasterBand(1)
    out_band.SetNoDataValue(np.nan)

    print("Creating NaN mask...")
    nan_mask = np.isnan(img_data)

    print("Replacing NaN with 0 for processing...")
    processing_data = np.nan_to_num(img_data, nan=0.0)

    print("Calculating global image variance (from original data)...")
    global_variance = np.nanvar(img_data)
    if global_variance == 0:
        global_variance = 1e-9
    print(f"Global variance: {global_variance}")

    print("Starting block processing...")
    for y in range(0, full_height, block_size):
        for x in range(0, full_width, block_size):
            block_height = min(block_size, full_height - y)
            block_width = min(block_size, full_width - x)

            filtered_block = process_block(processing_data, y, x, block_height, block_width, window_size, global_variance)
            mask_block = nan_mask[y:y + block_height, x:x + block_width]
            filtered_block[mask_block] = np.nan

            out_band.WriteArray(filtered_block, xoff=x, yoff=y)

    print("Block processing finished.")
    out_band.FlushCache()
    out_dataset.FlushCache()
    out_band = None
    out_dataset = None
    print(f"Filtered image saved to {output_path}")

if __name__ == '__main__':
    # Update these paths to your local files
    testfile = r'F:\s1a-iw-grd-vv.tiff'
    outputfile = r'D:\lee_filtered_s1a-iw-grd-vv.tif'

    try:
        print(f"Reading input file: {testfile}")
        img_data, dataset, geotransform, projection, gcps, gcp_projection = read_img(testfile)
        print(f"Geotransform: {geotransform}")
        print(f"Projection: {projection}")
        if gcps:
            print(f"Found {len(gcps)} GCPs in input image.")
        else:
            print("No GCPs found in input image.")
        block_processing(img_data, outputfile, dataset, geotransform, projection, gcps, gcp_projection, block_size=512, window_size=7)
    except FileNotFoundError as e:
        print(f"Error: {e}. Please ensure the input file path is correct.")
    except Exception as e:
        print(f"An unexpected error occurred: {e}")
    finally:
        dataset = None