1import tensorflow as tf
2from tensorflow.keras.saving import register_keras_serializable
3from tensorflow.keras import layers, models, backend as K
4import numpy as np
5
6@register_keras_serializable()
7def cold_temp_penalty(inputs):
8 temp = inputs[:, 0]
9 penalty = tf.where(
10 temp > 295.0,
11 1.0,
12 tf.where(
13 temp < 290.0,
14 0.0,
15 (temp - 290.0) / 5.0
16 )
17 )
18 return penalty[:, None]
19
20@register_keras_serializable()
21def fire_risk_booster(inputs):
22 temp = inputs[:, 0]
23 humidity = inputs[:, 1]
24 wind = inputs[:, 2]
25 veg = inputs[:, 3]
26
27 # Boost ranges
28 temp_boost = tf.sigmoid((temp - 305.0) * 1.2)
29 humidity_boost = tf.sigmoid((20.0 - humidity) * 0.5)
30 wind_boost = tf.sigmoid((wind - 15.0) * 0.8)
31 veg_boost = tf.sigmoid((veg - 70.0) * 0.5)
32
33 # Combine and scale
34 combined = temp_boost * humidity_boost * wind_boost * veg_boost
35 boost = 1.0 + 0.3 * combined # Up to 30% increase in fire score
36 return boost[:, None]
37
38@register_keras_serializable()
39def fire_suppression_mask(inputs):
40 temp = inputs[:, 0]
41 humidity = inputs[:, 1]
42 wind = inputs[:, 2]
43
44 # Suppress if warm but humid and still
45 temp_flag = tf.sigmoid((temp - 293.0) * 1.2)
46 humid_flag = tf.sigmoid((humidity - 50.0) * 0.4)
47 wind_flag = 1 - tf.sigmoid((wind - 5.0) * 0.8)
48
49 suppression = temp_flag * humid_flag * wind_flag
50 penalty = 1.0 - 0.3 * suppression # Max 30% suppression
51 return penalty[:, None]
