dataset = gdal.Open(file_name) # 打开文件

im_width = dataset.RasterXSize # 栅格矩阵的列数

im_height = dataset.RasterYSize # 栅格矩阵的行数

im_geotrans = dataset.GetGeoTransform() # 仿射矩阵

im_proj = dataset.GetProjection() # 地图投影信息

band = dataset.GetRasterBand(1)

im_array = band.ReadAsArray(0, 0, im_width, im_height) # 将数据写成数组，对应栅格矩阵

land_cover_names = []

for shp_file in shp_files:

basename = os.path.basename(shp_file) # 返回文件名

# split（'_'）函数将带有 '_' 的字符串分割成列表数据；然后 '_'.join（）函数将列表中的元素用'_'连接起来;

land_cover_name = '_'.join(basename.split('_')[:-1]) # 选取列表中除去最后一个元素的全部元素；

land_cover_names.append(land_cover_name) # 生成一个分类名列表

# 读取成dataFrame格式（将其看成series 组成的字典）；DataFrame的行索引是index，列索引是columns

train_data = pd.read_csv(csv_file)

xys = np.array(train_data[['xoff', 'yoff']]) # 获取图像坐标信息（xoff, yoff）

classes = np.array(train_data['class'], dtype=int) # 提取类别标签数据

del train_data # 关闭csv文件

cols, rows = zip(*xys) # zip()函数： 将列表中对应的元素打包成一个个元组

del xys

opt_samples = []

for opt_file in opt_files:

opt_ds = gdal.Open(opt_file)

bands = opt_ds.RasterCount

opt_sample = []

for i in range(1, bands+1, 1):

band = opt_ds.GetRasterBand(i)

band_min, band_max = band.ComputeRasterMinMax()

band_array = band.ReadAsArray() # ReadAsArray(始列， 始行，列间隔，行间隔)

data = band_array[rows, cols].flatten()

data = (data - band_min) * 255 / (band_max - band_min)

opt_sample.append(data)

del opt_ds

opt_list = np.vstack(opt_sample)

opt_samples.append(opt_list)

sar_samples = []

for sar_file in sar_files:

sar_ds = gdal.Open(sar_file)

# 维度为：波段数, 行数, 列数

sar_float32 = sar_ds.ReadAsArray()

sar_array = (sar_float32 * 255).astype(np.uint8)

sar_bandcount = sar_ds.RasterCount

# 将三维数据转化为二维数据并输出

sar_sample = sar_array[:, rows, cols].reshape(sar_bandcount, -1) # ( (rows * cols), 3)

sar_samples.append(sar_sample)

del sar_ds

samples = opt_samples + sar_samples

samples = np.vstack(tuple(samples)).T # 水平叠加延展 （cols* rows, 3*file_n）

samples = samples.astype(np.uint8)

print('smaples data is :', samples)

print('samples shape is:', samples.shape)

test_nan = np.sum(np.isnan(samples), axis=0)

print('test_nan_sum: ', test_nan.sum())

opt_ds = gdal.Open(opt_files[0])

rows = opt_ds.RasterYSize

cols = opt_ds.RasterXSize

opt_bands = opt_ds.RasterCount

sar_ds = gdal.Open(sar_files[0])

sar_bands = sar_ds.RasterCount

y_step = rows // batchs # “ // ” 表示整数除法，返回不大于结果的一个最大整数

y_residual = rows - (y_step * batchs)

del opt_ds, sar_ds

for i in range(0, batchs):

if i != batchs-1:

datas_opt = []

for opt_file in opt_files:

ds = gdal.Open(opt_file)

bands = ds.RasterCount

data_opt = []

for j in range(1, bands+1, 1):

band = ds.GetRasterBand(j)

band_min, band_max= band.ComputeRasterMinMax()

opt_array = band.ReadAsArray(0, i * y_step, cols, y_step) # ReadAsArray(始列， 始行，列间隔，行间隔)

data = (opt_array - band_min) * 255 / (band_max - band_min)

data = data.flatten()

data_opt.append(data)

del ds

opt_list = np.vstack(data_opt)

datas_opt.append(opt_list)

data_sar = []

for sar_file in sar_files:

ds = gdal.Open(sar_file)

data_float32 = ds.ReadAsArray(0, i * y_step, cols, y_step)

data = (data_float32 * 255).astype(np.uint8)

del ds

data = data.reshape(sar_bands, -1)

data_sar.append(data)

datas = datas_opt + data_sar

datas = np.vstack(datas)

else:

datas_opt = []

for opt_file in opt_files:

ds = gdal.Open(opt_file)

bands = ds.RasterCount

data_opt = []

for j in range(1, bands+1, 1):

band = ds.GetRasterBand(j)

band_min, band_max = band.ComputeRasterMinMax()

data = band.ReadAsArray(0, i * y_step, cols, y_step + y_residual) # ReadAsArray(始列， 始行，列间隔，行间隔)

data = (data - band_min) * 255 / (band_max - band_min)

data = data.flatten()

data_opt.append(data)

del ds

opt_list = np.vstack(data_opt)

datas_opt.append(opt_list)

data_sar = []

for sar_file in sar_files:

ds = gdal.Open(sar_file)

data_float32 = ds.ReadAsArray(0, i * y_step, cols, y_step + y_residual)

data = (data_float32 * 255).astype(np.uint8)

del ds

data = data.reshape(sar_bands, -1)

data_sar.append(data)

datas = datas_opt + data_sar

datas = np.vstack(datas)

RF_clf = RandomForestClassifier(n_estimators=150, max_depth=10, min_samples_split=100,

min_samples_leaf=100, n_jobs=-1)

# 再度分割数据集, 取1/3数据用作训练集

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.66, random_state=42)

# 拟合模型

RF_clf.fit(X_train, y_train)

# 做预测

y_pred = RF_clf.predict(X_test)

print("Accuracy: {0:.3f}\n".format(accuracy_score(y_test, y_pred)))

print('Kappa:', skll.kappa(y_test, y_pred))

print("\nConfusion matrix:\n{}\n".format(confusion_matrix(y_test, y_pred)))

labels = np.unique(np.concatenate((y_test, y_pred)))

scores_image = mglearn.tools.heatmap(

confusion_matrix(y_test, y_pred), xlabel='Predicted label',

ylabel='True label', xticklabels=labels, yticklabels=labels,

cmap='viridis', fmt='%d')

plt.title("Confusion matrix")

plt.gca().invert_yaxis()

print(classification_report(y_test, y_pred))

print("The training of the model ends!")

y_preds = []

for data in data_generator(opt_files, sar_files, batchs):

X = data

y_pred = RF_model.predict(X) # 预测结果 y_pred 输出为为行向量

y_pred[np.sum(X, 1) == 0] = 0 # 按列相加 -> 各波段都没有数据的像素设置为0

y_preds.append(y_pred)

X = np.array(X)

print('X is :{}\nX_shape is :{}'.format(X, X.shape))

pred = np.hstack(tuple(y_preds))

print('y = :{}\ny_shape = :{}'.format(pred, pred.shape))

x_size = in_ds.RasterXSize

y_size = in_ds.RasterYSize

y_array = y_pred.reshape((y_size, x_size))

predict_ds = make_raster(in_ds, out_fn, y_array, gdal.GDT_Byte, 0)

predict_ds.FlushCache()

levels = compute_overview_levels(predict_ds.GetRasterBand(1))

predict_ds.BuildOverviews('NEAREST', levels)