神经网络

Hi，大家好，我是半亩花海。在现代农业和环境监测中，了解土壤湿度的变化对于作物生长和水资源管理至关重要。通过深度学习技术，特别是卷积神经网络，我们可以利用过去的土壤湿度数据来预测未来的湿度趋势。本文将使用 PaddlePaddle 作为深度学习框架，通过数据分析、可视化、数据预处理、模型组网、模型训练和模型预测，基于卷积神经网络（CNN）模型来来处理时间序列数据，完成 10cm 土壤湿度的预测，从而实现一个简单的回归模型。

目录

一、导入必要库

二、数据分析

三、数据预处理

四、模型组网

五、模型训练

六、模型预测

七、完整代码

一、导入必要库

import time

import warnings

import numpy as np

import paddle

import paddle.nn as nn

import pandas as pd

import seaborn as sns

from matplotlib import pyplot as plt

from sklearn.preprocessing import MinMaxScaler

warnings.filterwarnings("ignore")

plt.rcParams['font.sans-serif'] = ['SimHei']

plt.rcParams['axes.unicode_minus'] = False

二、数据分析

soil_humidity = pd.read_excel("./soil_humidity.xlsx", engine="openpyxl")

soil_humidity["Datetime"] = pd.to_datetime(soil_humidity["datetime"])

soil_humidity.drop(["datetime"], axis=1, inplace=True)

soil_humidity.index = soil_humidity.Datetime

soil_humidity.drop(["Datetime"], axis=1, inplace=True)

soil_humidity = soil_humidity.sort_index()

print(soil_humidity.head())

sns.set(font='SimHei')

plt.figure(figsize=(8, 5))

plt.plot(soil_humidity.index, soil_humidity["10cm湿度(kg/m2)"], "b--", label='10cm湿度(kg/m2)')

plt.title("土壤湿度随时间变化关系", fontsize=14)

plt.xlabel("时间", fontsize=12)

plt.ylabel("10cm湿度(kg/m2)", fontsize=12)

plt.yticks(fontsize=12)

plt.xticks(fontsize=12)

plt.legend()

plt.grid(True, linestyle='--', alpha=0.5)

plt.show()

soil_humidity_10cm = soil_humidity.loc[soil_humidity.index[:], ['10cm湿度(kg/m2)']]

print(soil_humidity_10cm)

sns.set(font='SimHei')

corr = soil_humidity.corr()

plt.figure(figsize=(12, 8), dpi=100)

plt.title("数据框中各列之间的相关性", fontsize=13)

heatmap = sns.heatmap(corr, square=True, linewidths=0.2, annot=True, annot_kws={'size': 9})

heatmap.set_xticklabels(heatmap.get_xticklabels(), rotation=35, horizontalalignment='right')

heatmap.tick_params(axis='x', labelsize=8.5)

heatmap.tick_params(axis='y', labelsize=9)

cbar = heatmap.collections[0].colorbar

cbar.ax.tick_params(labelsize=9)

plt.show()

soil_humidity.head() 输出结果：

            10cm湿度(kg/m2)  40cm湿度(kg/m2)  ...  最大单日降水量(mm)  降水天数
Datetime                                  ...                   
2012-01-01          13.73          30.87  ...         0.51     5
2012-02-01          13.00          30.87  ...         0.76     5
2012-03-01          12.60          30.87  ...         4.83    13
2012-04-01          11.97          30.73  ...         5.33     3
2012-05-01          14.18          29.99  ...        15.49    10

[5 rows x 14 columns]

三、数据预处理

all_data = soil_humidity_10cm.values

split_fraction = 0.8

train_split = int(split_fraction * int(soil_humidity_10cm.shape[0]))

train_data = all_data[:train_split, :]

test_data = all_data[train_split:, :]

plt.figure(figsize=(8, 5))

plt.plot(np.arange(train_data.shape[0]), train_data[:, 0], label='train data')

plt.plot(np.arange(train_data.shape[0], train_data.shape[0] + test_data.shape[0]), test_data[:, 0], label='test data')

plt.title("数据集可视化", fontsize=14)

plt.xlabel("时间", fontsize=12)

plt.ylabel("10cm湿度(kg/m2)", fontsize=12)

plt.legend()

plt.show()

scaler = MinMaxScaler(feature_range=(-1, 1))

train_scal = scaler.fit_transform(train_data.reshape(-1, 1))

test_scal = scaler.fit_transform(test_data.reshape(-1, 1))

window_size = 12

train_scal = train_scal.reshape(-1)

train_scal = paddle.to_tensor(train_scal, dtype='float32')

def input_data(seq, ws):

out = []

L = len(seq)

for i in range(L - ws):

window = seq[i:i + ws]

label = seq[i + ws:i + ws + 1]

out.append((window, label))

return out

train_scal_data = input_data(train_scal, window_size)

print(train_scal_data[0])

train_scal_data[0] 这一组数据集的打印结果：

            10cm湿度(kg/m2)
Datetime                 
2012-01-01          13.73
2012-02-01          13.00
2012-03-01          12.60
2012-04-01          11.97
2012-05-01          14.18
...                   ...
2021-11-01          13.91
2021-12-01          13.14
2022-01-01          12.45
2022-02-01          12.10
2022-03-01          14.96

[123 rows x 1 columns]

四、模型组网

一维卷积层（convolution1d layer），根据输入、卷积核、步长（stride）、填充（padding）、空洞大小（dilations）一组参数计算输出特征层大小。

网络构造大体如下：

class CNNnetwork(paddle.nn.Layer):

def __init__(self):

super().__init__()

self.conv1d = paddle.nn.Conv1D(1, 1, kernel_size=2)

self.relu = paddle.nn.ReLU()

self.Linear1 = paddle.nn.Linear(11, 50)

self.Linear2 = paddle.nn.Linear(50, 1)

def forward(self, x):

x = self.conv1d(x)

x = self.relu(x)

x = self.Linear1(x)

x = self.relu(x)

x = self.Linear2(x)

return x

五、模型训练

paddle.seed(666)

model = CNNnetwork()

criterion = nn.MSELoss()

optimizer = paddle.optimizer.Adam(parameters=model.parameters(), learning_rate=0.001)

epochs = 30

split_idx = int(len(train_scal_data) * 0.8)

train_set = train_scal_data[:split_idx]

val_set = train_scal_data[split_idx:]

model.train()

start_time = time.time()

train_losses = []

val_losses = []

for epoch in range(epochs):

model.train()

train_loss = 0.0

for seq, y_train in train_set:

optimizer.clear_grad()

seq = paddle.reshape(seq, [1, 1, -1])

seq = paddle.to_tensor(seq, dtype='float32')

y_pred = model(seq)

y_train = paddle.to_tensor(y_train, dtype='float32')

loss = criterion(y_pred, y_train)

loss.backward()

optimizer.step()

train_loss += loss.numpy()[0]

model.eval()

val_loss = 0.0

with paddle.no_grad():

for seq_val, y_val in val_set:

seq_val = paddle.reshape(seq_val, [1, 1, -1])

seq_val = paddle.to_tensor(seq_val, dtype='float32')

y_val = paddle.to_tensor(y_val, dtype='float32')

val_pred = model(seq_val)

val_loss += criterion(val_pred, y_val).numpy()[0]

avg_train_loss = train_loss / len(train_set)

avg_val_loss = val_loss / len(val_set)

train_losses.append(avg_train_loss)

val_losses.append(avg_val_loss)

print('Epoch {}/{} - Train Loss: {:.4f} - Val Loss: {:.4f}'.format(epoch + 1, epochs, avg_train_loss, avg_val_loss))

print('nDuration: {:.0f} seconds'.format(time.time() - start_time))

plt.figure(figsize=(8, 5))

plt.plot(range(1, epochs + 1), train_losses, label='Train Loss')

plt.plot(range(1, epochs + 1), val_losses, label='Val Loss')

plt.title('Training and Validation Loss')

plt.xlabel('Epochs')

plt.ylabel('CNN_Loss')

plt.legend()

plt.show()

六、模型预测

将数据按 window_size 一组分段，每次输入一段后，会输出一个预测的值 y_pred，y_pred 与每段之后的第 window_size + 1 个数据作为对比值，用于计算损失函数。

例如前 5 个数据为 (1,2,3,4,5)，取前 4 个进行 CNN 预测，得出的值与 (5) 比较计算 loss。这里使用每组 13 个数据，最后一个数据作评估值，即 window_size=12。

"""

将数据按window_size一组分段，每次输入一段后，会输出一个预测的值y_pred

y_pred与每段之后的window_size+1个数据作为对比值，用于计算损失函数

例如前5个数据为(1,2,3,4,5),取前4个进行CNN预测,得出的值与(5)比较计算loss

这里使用每组13个数据,最后一个数据作评估值,即window_size=12

"""

preds = train_scal_data[-window_size:]

y_pred1 = []

model.eval()

for seq, y_train in preds:

seq = paddle.reshape(seq, [1, 1, -1])

seq = paddle.to_tensor(seq, dtype='float32')

result = model(seq)

y_pred1.append(result)

print("当前预测值：", y_pred1)

y_pred1 = np.array(y_pred1)

y_pred1 = y_pred1.reshape(-1, 1)

print("完整预测值：", y_pred1)

true_predictions = scaler.inverse_transform(y_pred1).reshape(-1, 1)

sns.set(font='SimHei')

plt.figure(figsize=(8, 5))

plt.plot(train_data[-window_size:], label='true_value')

plt.plot(true_predictions, label='predicted_value')

plt.title("真实值和预测值对比结果", fontsize=14)

plt.xlabel("最后12个值", fontsize=12)

plt.ylabel("10cm湿度(kg/m2)", fontsize=12)

plt.yticks(fontsize=12)

plt.xticks(fontsize=12)

plt.grid(True)

plt.legend()

plt.show()

完整预测值：

[[-0.8811799 ]
 [-0.31046718]
 [-0.09406683]
 [ 0.29082218]
 [ 0.64678204]
 [ 0.4292445 ]
 [ 0.11846957]
 [-0.17343275]
 [-0.36173454]
 [-0.55860955]
 [-0.6944711 ]
 [-0.6295543 ]]

七、完整代码

import time

import warnings

import numpy as np

import paddle

import paddle.nn as nn

import pandas as pd

import seaborn as sns

from matplotlib import pyplot as plt

from sklearn.preprocessing import MinMaxScaler

warnings.filterwarnings("ignore")

plt.rcParams['font.sans-serif'] = ['SimHei']

plt.rcParams['axes.unicode_minus'] = False

soil_humidity = pd.read_excel("./soil_humidity.xlsx", engine="openpyxl")

soil_humidity["Datetime"] = pd.to_datetime(soil_humidity["datetime"])

soil_humidity.drop(["datetime"], axis=1, inplace=True)

soil_humidity.index = soil_humidity.Datetime

soil_humidity.drop(["Datetime"], axis=1, inplace=True)

soil_humidity = soil_humidity.sort_index()

print(soil_humidity.head())

sns.set(font='SimHei')

plt.figure(figsize=(8, 5))

plt.plot(soil_humidity.index, soil_humidity["10cm湿度(kg/m2)"], "b--", label='10cm湿度(kg/m2)')

plt.title("土壤湿度随时间变化关系", fontsize=14)

plt.xlabel("时间", fontsize=12)

plt.ylabel("10cm湿度(kg/m2)", fontsize=12)

plt.yticks(fontsize=12)

plt.xticks(fontsize=12)

plt.legend()

plt.grid(True, linestyle='--', alpha=0.5)

plt.show()

soil_humidity_10cm = soil_humidity.loc[soil_humidity.index[:], ['10cm湿度(kg/m2)']]

print(soil_humidity_10cm)

sns.set(font='SimHei')

corr = soil_humidity.corr()

plt.figure(figsize=(12, 8), dpi=100)

plt.title("数据框中各列之间的相关性", fontsize=13)

heatmap = sns.heatmap(corr, square=True, linewidths=0.2, annot=True, annot_kws={'size': 9})

heatmap.set_xticklabels(heatmap.get_xticklabels(), rotation=35, horizontalalignment='right')

heatmap.tick_params(axis='x', labelsize=8.5)

heatmap.tick_params(axis='y', labelsize=9)

cbar = heatmap.collections[0].colorbar

cbar.ax.tick_params(labelsize=9)

plt.show()

all_data = soil_humidity_10cm.values

split_fraction = 0.8

train_split = int(split_fraction * int(soil_humidity_10cm.shape[0]))

train_data = all_data[:train_split, :]

test_data = all_data[train_split:, :]

plt.figure(figsize=(8, 5))

plt.plot(np.arange(train_data.shape[0]), train_data[:, 0], label='train data')

plt.plot(np.arange(train_data.shape[0], train_data.shape[0] + test_data.shape[0]), test_data[:, 0], label='test data')

plt.title("数据集可视化", fontsize=14)

plt.xlabel("时间", fontsize=12)

plt.ylabel("10cm湿度(kg/m2)", fontsize=12)

plt.legend()

plt.show()

scaler = MinMaxScaler(feature_range=(-1, 1))

train_scal = scaler.fit_transform(train_data.reshape(-1, 1))

test_scal = scaler.fit_transform(test_data.reshape(-1, 1))

window_size = 12

train_scal = train_scal.reshape(-1)

train_scal = paddle.to_tensor(train_scal, dtype='float32')

def input_data(seq, ws):

out = []

L = len(seq)

for i in range(L - ws):

window = seq[i:i + ws]

label = seq[i + ws:i + ws + 1]

out.append((window, label))

return out

train_scal_data = input_data(train_scal, window_size)

print(train_scal_data[0])

"""

一维卷积层（convolution1d layer），根据输入、卷积核、步长（stride）、填充（padding）、空洞大小（dilations）一组参数计算输出特征层大小。

网络大体构造如下：

先经过一维卷积层Conv1D。

使用ReLU激活函数对其进行激活。

然后经过第1层线性层Linear1。

再经过第2层线性层Linear2。

"""

class CNNnetwork(paddle.nn.Layer):

def __init__(self):

super().__init__()

self.conv1d = paddle.nn.Conv1D(1, 1, kernel_size=2)

self.relu = paddle.nn.ReLU()

self.Linear1 = paddle.nn.Linear(11, 50)

self.Linear2 = paddle.nn.Linear(50, 1)

def forward(self, x):

x = self.conv1d(x)

x = self.relu(x)

x = self.Linear1(x)

x = self.relu(x)

x = self.Linear2(x)

return x

paddle.seed(666)

model = CNNnetwork()

criterion = nn.MSELoss()

optimizer = paddle.optimizer.Adam(parameters=model.parameters(), learning_rate=0.001)

epochs = 30

split_idx = int(len(train_scal_data) * 0.8)

train_set = train_scal_data[:split_idx]

val_set = train_scal_data[split_idx:]

model.train()

start_time = time.time()

train_losses = []

val_losses = []

for epoch in range(epochs):

model.train()

train_loss = 0.0

for seq, y_train in train_set:

optimizer.clear_grad()

seq = paddle.reshape(seq, [1, 1, -1])

seq = paddle.to_tensor(seq, dtype='float32')

y_pred = model(seq)

y_train = paddle.to_tensor(y_train, dtype='float32')

loss = criterion(y_pred, y_train)

loss.backward()

optimizer.step()

train_loss += loss.numpy()[0]

model.eval()

val_loss = 0.0

with paddle.no_grad():

for seq_val, y_val in val_set:

seq_val = paddle.reshape(seq_val, [1, 1, -1])

seq_val = paddle.to_tensor(seq_val, dtype='float32')

y_val = paddle.to_tensor(y_val, dtype='float32')

val_pred = model(seq_val)

val_loss += criterion(val_pred, y_val).numpy()[0]

avg_train_loss = train_loss / len(train_set)

avg_val_loss = val_loss / len(val_set)

train_losses.append(avg_train_loss)

val_losses.append(avg_val_loss)

print('Epoch {}/{} - Train Loss: {:.4f} - Val Loss: {:.4f}'.format(epoch + 1, epochs, avg_train_loss, avg_val_loss))

print('nDuration: {:.0f} seconds'.format(time.time() - start_time))

plt.figure(figsize=(8, 5))

plt.plot(range(1, epochs + 1), train_losses, label='Train Loss')

plt.plot(range(1, epochs + 1), val_losses, label='Val Loss')

plt.title('Training and Validation Loss')

plt.xlabel('Epochs')

plt.ylabel('CNN_Loss')

plt.legend()

plt.show()

"""

将数据按window_size一组分段，每次输入一段后，会输出一个预测的值y_pred

y_pred与每段之后的window_size+1个数据作为对比值，用于计算损失函数

例如前5个数据为(1,2,3,4,5),取前4个进行CNN预测,得出的值与(5)比较计算loss

这里使用每组13个数据,最后一个数据作评估值,即window_size=12

"""

preds = train_scal_data[-window_size:]

y_pred1 = []

model.eval()

for seq, y_train in preds:

seq = paddle.reshape(seq, [1, 1, -1])

seq = paddle.to_tensor(seq, dtype='float32')

result = model(seq)

y_pred1.append(result)

print("当前预测值：", y_pred1)

y_pred1 = np.array(y_pred1)

y_pred1 = y_pred1.reshape(-1, 1)

print("完整预测值：", y_pred1)

true_predictions = scaler.inverse_transform(y_pred1).reshape(-1, 1)

sns.set(font='SimHei')

plt.figure(figsize=(8, 5))

plt.plot(train_data[-window_size:], label='true_value')

plt.plot(true_predictions, label='predicted_value')

plt.title("真实值和预测值对比结果", fontsize=14)

plt.xlabel("最后12个值", fontsize=12)

plt.ylabel("10cm湿度(kg/m2)", fontsize=12)

plt.yticks(fontsize=12)

plt.xticks(fontsize=12)

plt.grid(True)

plt.legend()

plt.show()

数据集可以评论区或者私信我哦。

相关知识

鸢尾花分类——神经网络详解
卷积神经网络的算法范文
“花朵分类“ 手把手搭建【卷积神经网络】
CNN卷积神经网络：花卉分类
应用卷积神经网络识别花卉及其病症
基于递归神经网络算法的电子物流配送系统配送路径优化
神经网络在病虫害诊断中的应用
基于matlab的神经网络的农业病虫害损失预测
深度学习之基于Tensorflow卷积神经网络花卉识别系统
用BP神经网络预测虫害

网址: 神经网络 https://www.huajiangbk.com/newsview169884.html