# load the original image X, y = loadData() height = y.shape[0] width = y.shape[1] PATCH_SIZE = 25 numComponents = 4 outputs = [] for i in range(height): row_outputs = [] for j in range(width): target = int(y[i, j]) if target == 0: row_outputs.append(0) else: # extract current patch and its label patch = X[i:i+PATCH_SIZE, j:j+PATCH_SIZE] target_patch = y[i:i+PATCH_SIZE, j:j+PATCH_SIZE] # convert patch into the suitable format to feed into model patch_tensor = patch.reshape(1, PATCH_SIZE, PATCH_SIZE, numComponents, 1).astype('float32') # make prediction for the patch using the trained model prediction = model.predict(patch_tensor) prediction = np.argmax(prediction, axis=1) row_outputs.append(prediction + 1) outputs.append(row_outputs) outputs = np.array(outputs) ground_truth = spectral.imshow(classes=y, figsize=(7, 7)) predict_image = spectral.imshow(classes=outputs.astype(int), figsize=(7, 7)) spectral.save_rgb("predictions.jpg", outputs.astype(int), colors=spectral.spy_colors)

这段代码是一个基于图像分类的模型的预测过程。首先从数据集中读取原始图像和对应的标签，然后通过循环遍历每一个像素点，提取当前像素点周围的 PATCH_SIZE x PATCH_SIZE 的图像块作为输入，将图像块转化为模型可以处理的格式，使用训练好的模型进行预测，将预测结果存储在一个矩阵中。最后使用 spectral 库的函数将预测结果可视化保存为一张图片。其中，代码注释中提到的 numComponents 是指图像中的波段数，输出结果中的 0 表示该像素点为背景，1~N 表示该像素点属于某个类别。