Data integration of humidity sensor and image texture for water content prediction of Gracilaria sp. during sun drying

Water content on site-measurement of dried seaweed required method with a minimum time of sample preparation time, less destructive effect to the sample, and could be validated. This research aimed to evaluate the potency of some features consist of image texture, resistance, and capacitance data of humidity sensor to predict water content changing of seaweed Gracilaria sp. during sun-drying. Dried Gracilaria sp. was rehydrated before being used in sun-drying for 4 hours. Gravimetrically-based water content evaluation, digital image taking, and measurement of resistance and capacitance value were conducted every 30 minutes interval during the drying. Images captured and collected by webcam in a conditioned lighting chamber were used subsequently for extraction of image texture features while a humidity sensor array contained 2 resistive sensors and 1 capacitive sensor respectively were applied to collect resistance and capacitance data. Collected data were used to create 4 datasets i.e. (1) 54 image texture features; (2) 3 resistance and capacitance features; (3) 57 features combination of dataset 1 and 2; and (4) 11 Features selected from dataset 3. Correlation coefficient and Root Mean Square Error of 4 datasets were applied for model evaluation utilized Multiple Linear Regression (MLR) and Multiple Layer Perceptron-based Neural Network (MLPNN). Investigation with cross-validation 10 folds test showed that MLPNN was the best model applied for dataset 1 with correlation coefficient and RMSE reached 0.89 and 9.11 respectively. Data integration of humidity sensor and image texture showed substantial potency to be used for the prediction of water content during sun drying.


Introduction
Indonesia has a large variety of seaweed species in its water. Not less than 782 seaweed species have been found. From the number of species, 18 species from5 genera have new market potency. Among the 5 genera, Eucheuma and Gracilaria were recently cultured [1].
The utilization of Gracilaria sp. in Indonesia is mainly used as a source of gelatin polysaccharide. Gracilaria sp. market in Indonesia has some significant issues i.e. low selling price, the competition level of other seaweed products, heterogeneity of its quality, and its environmental changing-related growth. Gracilaria sp. foreign market has strict quality requirements related to its growth rate and water content [2]. One method that can be applied to meet the food industrial standard of seaweed as an ingredient is a dehydration process [3]. Drying not only retard decay but also will prolong shelf life and helps to extract some chemical constituents [4]. Water content measurement of dried food ingredients is usually based on the oven drying method during a certain period to evaporate some water contained inside the food ingredient. The method was continued with a gravimetric method to measure the weight difference of food ingredients before and after oven drying. Despite the gravimetric method that has become reference was regulated in Indonesia National Standard [5] about dried seaweed but it needs sample preparation and drying time [6].
Seaweed farmers and distributors usually applied traditional inspection and handpicking during onsite dried seaweed water content measurement which requires extra time and labor [7]. Some researchers have developed a method to measure water content in food with faster, less or nondestructive, accurate, and more objective. Application of soil moisture sensor connected to microcontroller and internet of things to measure water content in fruit and grain with accuracy reached 86.7% has been investigated [6]while the utilization of image texture parameters altogether with shape and color to classify different phase of drying of apple slices was studied with accuracy reached 95% [8].
A capacitance type measuring device was designed and developed to estimate the water content of seaweed during sun drying. It was observed that there was a linear correlation between gravimetric and capacitance value with a correlation coefficient of 17.6% [9]. A resistance type measuring device was developed for measuring the water content of paddy rice, guinea corn, and millet with a negative coefficient of correlation -0.95, -0.99, and -0.99, respectively [10]. Capacitive and the resistive sensor were included as a semi-quantitative method for water content estimation which has an easiness to operate and maintain, good accuracy and response time, and easiness of power supply accessibility [11].
The determination method of water content in algae was conducted by applying thermogravimetric analysis to perform precision results [12]. An obstacle with the method is the balance precision has a positive correlation with its price [11].
The non-destructively estimation method of seaweed water content during drying is still limited to be studied. The humidity sensor and image texture features can be used as a quality indicator of food ingredients during drying because the data collection relatively easy, fast, and non-destructive. This research aimed to investigate the potency of the humidity sensor and image texture features to be used solely and or together in water content measurement of dried Gracilaria sp. during solar drying.

Sample preparation
Dried Gracilaria sp. were collected and bought from traders and distributors in the northern shore of Central Java and brought to the fish product processing laboratory of Indonesia Research Institute for Fisheries Postharvest Mechanization.
That seaweed was rehydrated by dipping it in fresh water for 30 minutes with seaweed water and the water ratio is 1:3(w/v). After being drained for 5 minutes that rehydrated seaweed was divided into three replications of 300 grams each. Those samples were ready to be used for solar drying with a time interval of 30 minutes for4hours.

Image acquisition and capacitance and resistance value measurement with humidity sensor
Image acquisition and capacitance and resistance value measurement with a humidity sensor for Gracilaria sp. were conducted with a time interval of 30 minutes during 4 hours of sun drying. Images were captured and collected with a webcam in the lighting chamber ( Figure 1 and Figure 2) and stored in.jpeg format with a resolution size 800x600 pixel. For image texture extraction, that .jpeg were converted and resized into .bmp format with size 400x300 pixel.
Mazda v 4.6 software was studied for image texture extraction [13]. The dried Gracilaria sp. image texture extraction process was depicted in Figure3. Capacitance and resistance value measurement with a humidity sensor was conducted by tapping the sensor tip over Gracilaria sp. until the read of value was stable on the instrument screen as shown in Figure 4. The instrument was supported for logging data into Micro SD Card and reset every time a new measurement was carried out.

Water content measurement based on SNI-01-2354-2-2006
Water content measurement was conducted as a validation method to data generated by image texture features and capacitive-resistive humidity sensors based on SNI-01-2354-2-2006 [14].  [15]. Model evaluation was tested with 10 folds cross-validation. The 10 folds CV split all data into 2 parts i.e. 10% as testing data while 90% remaining would be used as training data. The splitting process repeatedly to all data 10 times randomly. Correlation coefficient and Root Mean Square Error measured as average numbers from the repeated 10 times data split process [16].

Water content and drying rate changing during sun-drying
Changing water content during 240 minutes of sun drying showed a decreasing trend from 473.39 % dry basis (d.b) to 18.46 % d.b as showed in Figure 5. The main factor that contributes to this trend was solar irradiation as observed by [17].
Archive of climate data on 09 th June 2020 in the Bantul region showed average temperature and humidity, 26.5 o C, and 85%, respectively [18]. The instability of the drying rate as depicted in Figure 6 is very affected by the environmental climate during drying. It was observed that the drying rate could increase if the drying temperature was high while the relative humidity measured was low [19]. This phenomenon could happen when the partial pressure of water vapor and vapor pressure had a substantial difference.  Figure 6. Drying rate changes during drying.

Datasets
Datasets produced from image texture extraction had 54 features as mentioned in Table 1. Image texture features extracted with Mazda software could reach a high number and variability of parameters. To use the large features so conclusions could be formulated or ease to interpret analysis result number of features needed to be reduced or selected [20].  The humidity sensor used to collect resistance and capacitance value during Gracilaria sp. solar drying shows the results in Figures 7, 8, and 9. The resistance value of sensor X2 less sensitive than sensor X1 and X3 which performed zero values read after 180 minutes of drying. The use of 2 resistance value-based soil humidity sensors to predict the water content of fruit and grain was investigated and was able to reach an accuracy level of 86.7% [6]. Dataset 4 was generated by applying correlation-based feature selection (Cfs) in Weka v 3.8 Software. This method evaluated the feasibility of features subset by taking into account 2 parts i.e. predictive ability of each feature and redundancy level among those features [21].

Water content prediction model evaluation
The model evaluation showed in Table2 performed that 54 image texture features had the biggest potency to be used as a predictor in the water content predictive model of Gracilaria sp. during sun drying based on the highest correlation rate and lowest RMSE. It was investigated that variables easy to measure (color, texture, and spectral distribution) can be used to evaluate variables difficult to measure, such as chemical composition [21]. Furthermore, texture features were calculated based on pixel brightness information and mutual interaction between them. MLPNN showed a higher correlation for three datasets compared to MLR. MLPNN more represented non-linear boundary decisions [15].

Conclusions
It is observed from the experiment that dataset 1(image texture features) performed the best prediction of dried Gracilaria sp. water content using MLPNN modeling with a coefficient of correlation reached above 87% and RMSE 9.1136. Data integration followed by feature selection and MLR modeling (Dataset 4) performed promising results for the prediction of dried Gracilaria sp. water content. It has fewer features but its estimation ability almost equal to those of image texture features. Image texture features and humidity sensor can be used as a non-destructive method to predict Gracilaria sp. during drying. Future works for this research requires attempts to increase robustness in the model and to compare model effectiveness for other high-demanded seaweed species.