*4.1. General Characteristics*

The global trend of the audio database obtained from the sensor network has been firstly analyzed through focusing mainly in RTN, disregarding specific characteristics of the ANE observed during recordings. In this regard, spectrum-time profiles [54] defined as the hourly time evolution of the mean spectrum of the audio has been computed for each sensor and day, following the same approach as in [24]. Their computation has been performed using the 48 frequency sub-bands of the Mel-Frequency Cepstral Coefficients (MFCC) features of the incoming audio signal, following the setup explained in [55], and computing the mean spectrum along the 20 min of audio gathered every hour. This mean spectrum obtained for each of the 24 h in a day conforms the spectrum-time profile of a complete measured day, which mainly includes RTN raw signal.

Figure 5 shows four examples of the measured spectrum-time profiles, for two sensors hb134 and hb156 and for the two recorded days of the campaign. As can be observed, the general trend of the spectrum-time profile during weekday (Figure 5a,b) is quite similar for these two sensors, while sensor hb156 presents a slightly higher energy at high frequencies and during a narrower daily period. During the weekend (Figure 5c,d) a very persistent ANE of hard rain was present in almost all sensors, which can be seen as a high spectrum-profile value around 14 h. Additionally, another aspect that can be highlighted comparing week spectrum-time profiles with those of the weekend is that during weekend the raising and decreasing of acoustic energy during daily period is smoother than during the week day, which can be explained by the fact that on work days there are traffic rush hours at the beginning and ending of working hours.

(**a**) Spectrum-time profile of sensor hb134 at week recording day.

(**b**) Spectrum-time profile of sensor hb156 at week recording day.

(**c**) Spectrum-time profile of sensor hb134 at weekend recording day.

(**d**) Spectrum-time profile of sensor hb156 at weekend recording day.

**Figure 5.** Spectrum-time profiles of sensors hb134 and hb156 at weekday (the 2nd of November 2017) and at weekend day (the 5th of November 2017).

### *4.2. Overall Analysis*

Table 1 lists the distribution of the ANEs in terms of their number of occurrences observed within the recorded database, showing both the aggregated and segregated distributions for all sensors, respectively. The total amount of ANEs labeled in the dataset generation represents a 1.8% of all the recorded time, together with around 3.4% of raw data labeled as *cmplx*, which is the data result of the mixture of noises or with unidentifiable sound.


**Table 1.** Aggregated ANE occurrences and their total duration distribution per category.

Table 1 shows the distribution of total ANE duration segregated by ANE subcategory. Most frequently observed ANEs were rain (6413.5 s) and thunder (753.4 s), due do the thunderstorm episode observed during the weekend day in the city of Rome at 14:00. Non-meteorological ANEs with higher total duration values were sounds of vehicles brakes (1387.2 s), trains (655.7 s), birds (338.9 s), sirens (327.8 s), horns (165.7 s), and sounds of trucks (382.4 s). ANEs with mid-total duration values were alarm sounds (84.3 s), interfering sounds (69.4 s), structure movement sounds (45.8 s), noise of pressurized air or *busd* (56.6 s), and music (31.6 s), while ANEs with lowest presence during the recordings were door sounds (16.3 s), airplanes (14.5 s), and bikes (9.6 s).

In Figure 6, the boxplots of the ANE durations are shown. The longer ANE correspond to *musi* and *alrm*, while the shorter are the *door*, *bird* and *busd*.

**Figure 6.** Boxplots of duration for each ANE category. Logarithmic axis of duration is used for a better observation of the duration values.

To determine the salience of the ANEs, the SNR measure is calculated by following the steps in Section 3.3.1. Figure 7 shows the SNR distributions for each ANE category. It can be observed that ANEs with higher SNR are sounds of trains, bikes, sirens, trucks, and horns.

**Figure 7.** Boxplots of SNR for each ANE subcategory.
