Keras CNN predicts 2 classes out of 4 - tensorflow

Question

I have a problem about my CNN model made using tensorflow. The goal is to predict the classes of satellite images, corresponding to the type of clouds (data extracted from the kaggle competition "Planet: Understanding the Amazon from Space"). There are 4 classes : clear, cloudy, partly cloudy and haze.
Everything works fine until I try to test the model on individual images. Then, it always predicts 2 classes and nothing else. I noticed that if I run the model again, it may predict 2 other classes among the 4. The model was trained for 10 epochs, which gave an accuracy of 0.8717.
Here is my code :
import numpy as np
import pandas as pd
import cv2
from tqdm import tqdm
import h5py
import os
os.listdir("/kaggle/input/")
import tensorflow as tf
from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense,MaxPooling2D,Conv2D,Flatten,Dropout,Activation
from tensorflow.keras.layers import BatchNormalization
from sklearn import svm
from sklearn.model_selection import cross_val_score
import matplotlib.pyplot as plt
from tensorflow.keras.preprocessing.image import ImageDataGenerator, load_img
from oauth2client.client import GoogleCredentials
import csv
#from keras.optimizers import RMSprop
from tensorflow.keras import Input, Model
batch_size = 128
img_width = 256
img_height = 256
train_data = ImageDataGenerator(
rescale = 1./255,
validation_split = 0.25)
train_generator = train_data.flow_from_directory(
'../input/clouds',
target_size=(img_height, img_width),
color_mode='rgb',
batch_size=batch_size,
shuffle = True,
class_mode="categorical",
subset = 'training'
)
valid_generator = train_data.flow_from_directory(
'../input/clouds',
target_size=(img_height, img_width),
batch_size=batch_size,
class_mode='categorical',
subset = 'validation'
)
num_classes = 4
model = Sequential([
Input(shape = [img_width, img_height, 3]),
Conv2D(128,4,activation = 'relu'),
MaxPooling2D(),
Conv2D(64,4,activation = 'relu'),
MaxPooling2D(),
Conv2D(32,4, activation = 'relu'),
MaxPooling2D(),
Conv2D(16,4,activation = 'relu'),
MaxPooling2D(),
Flatten(),
Dense(64, activation = 'relu'),
Dense(num_classes, activation = 'softmax')
])
model.compile(optimizer = "adam",
loss = 'categorical_crossentropy',
metrics=['accuracy'])
model.fit(train_generator, validation_data = valid_generator, epochs = 10)
img_to_predict = cv2.imread('/kaggle/input/clouds-test/clouds_test/test_3877_6013089.jpg') #an augmented image from original dataset
img_to_predict = cv2.cvtColor(img_to_predict, cv2.COLOR_BGR2RGB)
img_to_predict = np.expand_dims(cv2.resize(img_to_predict, (256,256)), axis = 0)
res = model.predict(img_to_predict)
label_map = (train_generator.class_indices)
print(label_map)
print(list(label_map)[np.argmax(res, axis = -1)[0]])
Thank you for you help.