IoT-based control and monitoring system for hydroponic plant growth using image processing and mobile applications

Internet of Things system for controlling and monitoring the growth of hydroponic

The implementation of the Internet of Things (IoT) in a smart hydroponic system is designed to facilitate the monitoring and control of various devices or systems automatically and efficiently to support the growth of hydroponic plants. To create an accurate and precise system, a series of technical processes are required, from calibration to component integration. The following are the main components involved in the IoT implementation in HydroFarm, as illustrated in Fig. 2.

The wiring process of components, as shown in Fig. 2, is necessary to assist the builder in determining the number of wires and creating the wiring paths. The printed circuit board (PCB) design in Fig. 3 is intended to streamline the shape of the device, with the design adjusted according to the component wiring. This design enhances system compactness, minimizes signal interference, and improves electrical stability, thereby increasing overall reliability and performance. Ultimately, it streamlines hardware assembly while contributing to the efficiency and effectiveness of the IoT-based hydroponic system.

To build an IoT-based smart hydroponic monitoring and controlling system, several essential components are required:

Convolutional neural network for detecting plant health or conditions

In HydroFarm, image processing is employed to detect the condition of hydroponic plants, specifically focusing on the leaves. This process involves developing a model through training data and implementing a detection system in a mobile application. A common neural network architecture used for image analysis is the CNN, which simplifies images to provide more accurate analysis results (Desai & Shah, 2021). CNN consists of key layers such as convolutional layers, pooling layers, and fully connected layers, influenced by parameters like input, filter, padding, stride, pooling, weight, and bias (Mulyawan, 2025).

Using the VGG16 for leaf health detection on a small dataset can be justified due to its robust feature extraction capabilities. VGG16 comprises 13 convolutional layers, three fully connected layers, and five pooling layers, and is pre-trained on the ImageNet dataset. Its weights are fine-tuned and extended with custom layers to classify leaf conditions into two categories: healthy and unhealthy. By leveraging pretrained weights, VGG16 adapts well to small datasets without requiring extensive data or training time, reducing the risk of overfitting. Optimized for mobile applications, the model is converted to TensorFlow Lite format for efficient deployment (Yang et al., 2021). Its well-established performance and ease of implementation make it a reliable baseline model for benchmarking and further optimization, despite the added complexity.

The dataset used in this project includes images of pakchoi and water spinach, consisting of 371 images categorized into two classes, resized to 240 × 240 pixels to match CNN model’s input requirements and split into training, testing, and validation sets. To enhance generalization and reduce overfitting, data augmentation is applied using Image Data Generator with transformations including rescaling pixel values to range [0, 1], shear distortions, horizontal and vertical flips, brightness adjustment (50% to 150%), random rotation (up to 40 degrees), and zooming (20%). The training process is configured with batch size of 15, 50 epochs, and dynamically computed staps per epoch and validation steps based on dataset size. ADAM optimizer with learning rate 0.001 is used alongside categorical cross-entropy loss function, with ReLU activation for intermediate layers and Softmax for output layer to handle multi-class classification. After training, the model’s performance is evaluated based on validation accuracy with the best epoch recorded for further analysis.

In the implementation process, as shown in Fig. 4, the process begins with collecting a dataset of hydroponic plant leaf images categorized by their condition (healthy and unhealthy) over a 15-day period, which is divided into three sets with distribution ratios of 70% for training, 15% for testing, and 15% for validation in the first test; and 80% for training, 10% for testing, and 10% for validation in the second test.

Next, preprocessing is performed to enhance image quality, including rescaling (1/255), shear transformation (0.2), horizontal and vertical flips, brightness adjustment ([0.5, 1.5]), rotation (40), and zooming (0.2). After preprocessing, the training dataset is used to train the CNN model by adjusting parameters such as learning rate and batch size, followed by evaluating the model’s performance using the testing and validation datasets to ensure accuracy exceeds 85%. If the performance criteria are not met, analysis, debugging, and hyperparameter tuning are conducted before retraining the model. Once the criteria are met, the model is saved in .tflite format and is ready to be implemented for detecting hydroponic plant conditions.

Figure 5 illustrates the process of implementing the CNN model in HydroFarm, beginning with capturing or uploading an image of a hydroponic leaf. The image undergoes preprocessing, where it is resized to 240 × 240 pixels to match the required input shape for the CNN model. The preprocessed image is then fed into CNN model for analysis. If the model cannot detect a leaf in the input image, an error message is displayed, prompting the user to upload another image. If a leaf is successfully detected, the model proceeds with classifying the leaf’s condition. The classification results are then displayed to the user through the system interface. The CNN model is implemented using a Sequential architecture and optimized with the Adam optimizer. The dataset for training and evaluation includes primary data collected from hydroponic farmers camera, as well as secondary data sourced from Google, focusing on the analysis of green pigments to detect the condition of hydroponic plants via mobile application.

Mobile application as a user interface for monitoring and controlling IoT and CNN integration

The development of the HydroFarm application interface, as shown in Fig. 6, was conducted using Kotlin, the native programming language for Android application development. For this development process, Android Studio was utilized as the primary integrated development environment (IDE). Android Studio was chosen for its comprehensive support for Android application development, including features for debugging, user interface (UI) visualization, and device simulation through the Android Emulator.

The HydroFarm application is designed to integrate two main components: the deep learning (DL) model implemented using TensorFlow Lite (.tflite) and real-time sensor data received from Firebase Realtime Database. The DL model is used to detect the condition of plants based on the provided input, while Firebase serves as the central storage and management hub for sensor data from the IoT devices installed in the hydroponic system.

Results of the Internet of Things system

The following is an overview of the testing of the IoT components for the control and monitoring system of hydroponic plant growth.

ESP32-DevKitC V4

The ESP32 operates effectively as the central controller, processing sensor data and managing communication between sensors, actuators, and the cloud. The ESP32 is able to drive all the sensors and pumps simultaneously, as demonstrated by the successful testing of the other sensors/modules and the effective transmission of data to the Firebase realtime database, as shown in Fig. 8.

pH SEN0161-V2 sensor

The evaluation of water pH measurement involved integrating the pH sensor with a microcontroller, supplying power to the system, and programming the microcontroller to activate the sensor’s functionality. The device was positioned near the hydroponic plant to observe and analyze the water pH data provided by the sensor, assessing its accuracy and determining the need for calibration. If calibration was required, it was performed by reprogramming the system using a linear regression formula, followed by data analysis of the observations.

After conducting tests, the following are the results from the pH sensor in providing water pH information when the probe was immersed in calibration solutions of pH 4.00 (25 °C), pH 6.86 (25 °C), and pH 9.18 (25 °C), as shown in Tables 3–5.

Table 3:

pH sensor testing results using pH 4.00 solution before calibration.

No.	Calibration solution pH 4.00 (25 °C) Before calibration
	pH sensor	pH tester	Difference in pH value	Percentage error pH
1.	2.88	4.00	1.12	38.89%
2.	2.88	4.00	1.12	38.89%
3.	2.88	4.00	1.12	38.89%
4.	2.88	4.00	1.12	38.89%
5.	2.88	4.00	1.12	38.89%
Average	2.88	4.00	1.12	38.89%

DOI: 10.7717/peerj-cs.2763/table-3

Table 4:

pH sensor testing results using pH 6.86 solution before calibration.

No.	Calibration solution pH 6.86 (25 °C) Before calibration
	pH sensor	pH tester	Difference in pH value	Percentage error pH
1.	2.23	6.86	4.63	207.62%
2.	2.23	6.86	4.63	207.62%
3.	2.23	6.86	4.63	207.62%
4.	2.23	6.86	4.63	207.62%
5.	2.23	6.86	4.63	207.62%
Average	2.23	6.86	4.63	207.62%

DOI: 10.7717/peerj-cs.2763/table-4

Table 5:

pH sensor testing results using pH 9.18 solution before calibration.

No.	Calibration solution pH 9.18 (25 °C) Before calibration
	pH sensor	pH tester	Difference in pH value	Percentage error pH
1.	1.68	9.18	7.5	446.43%
2.	1.68	9.18	7.5	446.43%
3.	1.68	9.18	7.5	446.43%
4.	1.68	9.18	7.5	446.43%
5.	1.68	9.18	7.5	446.43%
Average	1.68	9.18	7.5	446.43%

DOI: 10.7717/peerj-cs.2763/table-5

From the results of the above tests, there is a significant discrepancy between the values measured by the pH sensor and the calibration solutions used as a comparison. Since the test results do not match the expected values of the solutions, calibration of the pH sensor coding is required. The calibration method used is a linear regression approach. This linear regression approach is necessary to minimize the difference between the pH sensor readings and the solution values. The formula obtained is shown in Fig. 9.

The testing was conducted again with the pH sensor that had been calibrated using the linear regression formula. Below are the results from the calibrated pH sensor testing while the probe was immersed in calibration solutions at pH 4.00 (25 °C), pH 6.86 (25 °C), and pH 9.18 (25 °C), as shown in Tables 6–8. From these tables, it is shown that the data remained consistent, indicating that the pH sensor is capable of providing consistent and sufficiently accurate information, as evidenced by its small percentage error.

Table 6:

pH sensor testing results using pH 4.00 solution after calibration.

No.	Calibration solution pH 4.00 (25 °C) After calibration
	pH sensor	pH tester	Difference in pH value	Percentage error pH
1.	4.10	4.00	0.1	2.44%
2.	4.10	4.00	0.1	2.44%
3.	4.10	4.00	0.1	2.44%
4.	4.10	4.00	0.1	2.44%
5.	4.10	4.00	0.1	2.44%
Average	4.10	4.00	0.1	2.44%

DOI: 10.7717/peerj-cs.2763/table-6

Table 7:

pH sensor testing results using pH 6.86 solution after calibration.

No.	Calibration solution pH 6.86 (25 °C) After calibration
	pH sensor	pH tester	Difference in pH value	Percentage error pH
1.	6.90	6.86	0.04	0.58%
2.	6.90	6.86	0.04	0.58%
3.	6.90	6.86	0.04	0.58%
4.	6.90	6.86	0.04	0.58%
5.	6.90	6.86	0.04	0.58%
Average	6.90	6.86	0.04	0.58%

DOI: 10.7717/peerj-cs.2763/table-7

Table 8:

pH sensor testing results using pH 9.18 solution after calibration.

No.	Calibration solution pH 9.18 (25 °C) Before calibration
	pH sensor	pH tester	Difference in pH value	Percentage error pH
1.	9.20	9.18	0.02	0.217%
2.	9.20	9.18	0.02	0.217%
3.	9.20	9.18	0.02	0.217%
4.	9.20	9.18	0.02	0.217%
5.	9.20	9.18	0.02	0.217%
Average	9.20	9.18	0.02	0.217%

DOI: 10.7717/peerj-cs.2763/table-8

TDS SEN0244 sensor

The evaluation of TDS measurement involved integrating the TDS sensor with a microcontroller, supplying power to the system, and programming the microcontroller to activate the sensor’s functionality. The device was positioned near the hydroponic plant to observe and analyze the TDS data provided by the sensor, assessing its accuracy and determining the need for calibration. If calibration was required, it was performed by reprogramming the system using a linear regression formula, followed by data analysis of the observations.

After conducting the tests, the following are the results from the TDS sensor in providing total dissolved solids information when the probe was immersed in calibration solutions of 342 ppm (25 °C) and 500 ppm (25 °C), as shown in Tables 9 and 10.

Table 9:

TDS sensor testing results using 342 ppm calibration solution (25 °C) before calibration.

No.	Calibration solution 342 ppm (25 °C) Before calibration
	TDS sensor	TDS tester	Difference in TDS value	Percentage error TDS
1.	138	342	204	147.826%
2.	138	342	204	147.826%
3.	138	342	204	147.826%
4.	138	342	204	147.826%
5.	138	342	204	147.826%
Average	138	342	204	147.826%

DOI: 10.7717/peerj-cs.2763/table-9

Table 10:

TDS sensor testing results using 500 ppm calibration solution (25 °C) before calibration.

No.	Calibration solution 500 ppm (25 °C) Before calibration
	TDS sensor	TDS yester	Difference in TDS value	Percentage error TDS
1.	195	500	305	156.41%
2.	195	500	305	156.41%
3.	195	500	305	156.41%
4.	195	500	305	156.41%
5.	195	500	305	156.41%
Average	195	500	305	156.41%

DOI: 10.7717/peerj-cs.2763/table-10

From the test results above, there is a significant difference between the values measured by the TDS sensor and the calibration solution values used for comparison. Since the test results did not align with the solution values, calibration of the TDS sensor coding is necessary. The calibration approach used is linear regression. This linear regression approach is essential to minimize the discrepancy between the TDS sensor values and the solution values. The formula obtained is shown in Fig. 10.

The testing was conducted again using the TDS sensor that had been calibrated with the linear regression formula. The results of the calibrated TDS sensor testing, while the probe was immersed in calibration solutions of 342 ppm (25 °C) and 500 ppm (25 °C), are shown in Tables 11 and 12. From these tables, it is shown that the data remained consistent, indicating that the TDS sensor is capable of providing consistent and sufficiently accurate information, as evidenced by its small percentage error.

Table 11:

TDS sensor testing results using 342 ppm calibration solution (25 °C) after calibration.

No.	Calibration solution 342 ppm (25 °C) After calibration
	TDS sensor	TDS tester	Difference in TDS value	Percentage error TDS
1.	345	342	3	0.87%
2.	345	342	3	0.87%
3.	345	342	3	0.87%
4.	345	342	3	0.87%
5.	345	342	3	0.87%
Average	345	342	3	0.87%

DOI: 10.7717/peerj-cs.2763/table-11

Table 12:

TDS sensor testing results using 500 ppm calibration solution (25 °C) after calibration.

No.	Calibration solution 500 ppm (25 °C) After calibration
	TDS sensor	TDS tester	Difference in TDS value	Percentage error TDS
1.	508	500	8	1.575%
2.	508	500	8	1.575%
3.	508	500	8	1.575%
4.	508	500	8	1.575%
5.	508	500	8	1.575%
Average	508	500	8	1.575%

DOI: 10.7717/peerj-cs.2763/table-12

LCD I2C 20 × 4

After implementation and testing, the LCD demonstrated good performance, as shown in Fig. 11, accurately and clearly displaying data on TDS, pH, water temperature, air temperature, and humidity. Users can easily monitor environmental conditions, allowing for the evaluation of water quality and timely actions to maintain ideal parameters for plant growth.

Information regarding water and air temperatures helps users identify changes that could affect plant growth, enabling quick adjustments to the temperature control system. Additionally, the real-time displayed humidity data provides crucial insights to prevent stress on the plants and maintain their overall health. Overall, this system supports more effective monitoring and management of the environment within the Smart Hydroponic system.

Results of convolutional neural network

Testing was conducted using two scenarios for splitting the data into training, testing, and validation sets: 70%, 15%, and 15% for the first test, and 80%, 10%, and 10% for the second test. In the first test, the resulting model achieved an accuracy of 96.1%, with precision, recall, and F1-score values shown in Fig. 12, along with a computation time of 10 s.

Next, a check was performed on the test dataset, and the classification results were mapped in the form of a confusion matrix. Figure 13 shows the confusion matrix for classifying hydroponic plant leaves, where out of 38 healthy leaf datasets, 36 were classified as healthy and two were classified as unhealthy. For the 39 unhealthy leaves, 38 were classified as unhealthy, while one dataset was incorrectly classified as healthy.

In addition to testing the data against the model, validation testing was conducted to ensure that the created model does not suffer from overfitting or underfitting by comparing the test accuracy results with the validation accuracy. Figure 14 shows that the first test has a test accuracy of 96.1% and a validation accuracy of 98.78%. The validation dataset accuracy is higher than the test dataset accuracy; however, the small difference indicates that the model performs well on the test dataset and does not exhibit significant overfitting.

In the second test, the resulting model achieved an accuracy of 90.38%, along with precision, recall, and F1-score values as shown in Fig. 15, with a computation time of 6 s.

Next, a check was performed on the test dataset, and the classification results were mapped in the form of a confusion matrix. Figure 16 shows the confusion matrix for the classification of hydroponic plant leaves, where out of 26 healthy leaf datasets, 23 were classified as healthy and three as unhealthy. Meanwhile, for the 26 unhealthy leaves, 24 were classified as unhealthy and two as healthy.

In addition to testing the model against the test dataset, validation data testing was conducted to ensure that the created model does not experience overfitting or underfitting by comparing the test accuracy results with the validation accuracy. Figure 17 shows that the first test had a test accuracy of 90.38% and a validation accuracy of 100%. Among these values, the accuracy on the validation dataset is higher than that on the test dataset. The validation dataset accuracy is very high, reaching 100%, indicating that the model fits the validation data exceptionally well. However, an accuracy of 100% on the validation dataset may indicate potential overfitting on a relatively small dataset.

By comparing both scenarios, the model from the first test is deemed better for subsequent testing. After conducting tests with the previously designed scenarios, the model can detect the condition of hydroponic plant leaves, accompanied by confidence values. For hydroponic plants with green leaf characteristics, the model detects them as healthy plants with an accuracy of 93.56%. For hydroponic plants with yellowing leaves, the model classifies them as unhealthy plants with an accuracy of 51.84%. Additionally, testing on wilted pakchoi plants resulted in the model detecting them as unhealthy plants with an accuracy of 99.62%. The testing was conducted directly using the application, as shown in Fig. 18.

Result of mobile application

In an effort to develop an IoT-based smart hydroponic system integrated with a Convolutional Neural Network (CNN) model for image processing, an effective and intuitive application is needed. This application is designed with three main menus: Home, Take Photo, and Profile.

The Home menu in the developed smart hydroponic application serves as the main interface, where users can monitor environmental data in real-time, including TDS, pH, water temperature, air temperature, and humidity enviroment. This visual information is designed to facilitate users in making decisions regarding the management of the hydroponic system. This menu is equipped with control features that allow users to manage the nutrient pump both manually and automatically, providing enhanced flexibility in adjusting nutrient levels based on the specific requirements of different plant varieties.

In the automatic control feature, as shown in Fig. 19, users can set the desired pH and TDS range; after entering these values, the system will automatically adjust the operation of the nutrient pump according to the established parameters. For example, if the pH or TDS values fall outside the range inputted by the user, the nutrient pump will be activated or deactivated automatically to maintain stable conditions for the plants, allowing the system to function autonomously and responsively to changes in the environment. Additionally, the application also provides a manual control option, as shown in Fig. 20, where users can press the provided button to turn the nutrient pump on or off directly according to their preferences. This feature offers added flexibility for users who want to take control and make adjustments based on their specific needs at that moment.

The Take Photo menu, as shown in Fig. 21, integrated within the HydroFarm application, provides users with the capability to capture plant images directly using their smartphone camera or select images from the gallery. This feature is designed to facilitate the collection of visual data, which is then processed by the CNN model embedded in the application. The use of this technology is crucial for detecting plant health conditions more accurately and efficiently.

Once the plant image is captured or uploaded, the CNN model analyzes the image to detect various health indicators of the plant. This CNN-based analysis enables early identification of potential issues, allowing users to take corrective actions promptly before further damage occurs.

The Profile menu, as shown in Fig. 22, in the developed smart hydroponic application, is designed to provide educational information about the HydroFarm app and guidelines for its usage. This feature aims to help users gain a deeper understanding of the app’s core concept and primary objectives, as well as how to operate it effectively for managing their hydroponic system.

To determine the feasibility of the HydroFarm application, the researcher used the system usability scale (SUS) method. SUS, created by John Brooke (Huda et al., 2023), is a measurement tool used to assess the functionality of a product by testing it directly with users. This method is beneficial for evaluating an application based on its features and usability from the user’s perspective.

The SUS method was carried out via a Google Form containing 10 questions rated on a five-point scale, ranging from “Strongly Disagree” to “Strongly Agree.” Specifically, there are five positive and five negative questions. The 10 questions are as follows:

1. I feel comfortable using HydroFarm.

2. I find the features in HydroFarm too complex to use.

3. I find HydroFarm easy to use.

4. I think I would need technical support to use HydroFarm effectively.

5. The features in HydroFarm are well-integrated.

6. I find too much inconsistency within HydroFarm.

7. I believe most people can learn to use HydroFarm quickly.

8. I find HydroFarm too confusing to use.

9. I feel confident using HydroFarm.

10. I need to learn a lot before I can effectively use HydroFarm.

The scoring will be based on user responses, and the results will be calculated using a formula specified in the reference (Huda et al., 2023). Each respondent’s SUS score will then be averaged across all responses using the formula (Welda, Putra & Dirgayusari, 2020).

(1) $$x_i=\left(\begin{array}{c}\left(Q_{1i}-1\right)+\left(5-Q_{2i}\right)+\left(Q_{3i}-1\right)+\\ \begin{array}{c}\left(5-Q_{4i}\right)+\left(Q_{5i}-1\right)+\left(5-Q_{6i}\right)+\\ \left(Q_{7i}-1\right)+\left(5-Q_{8i}\right)+\left(Q_{9i}-1\right)+\\ \left(5-Q_{10i}\right)\end{array}\end{array}\right).$$

Note:

x_i = The total score of each respondent.

$Q_{\left(1-10\right)i}=$ The score for each question (1 to 10) for respondent i

(2) $$\overline{x}={\displaystyle \frac{1}{N}\sum _{i=1}^N\left(x_i\times 2.5\right).}$$

Note:

$$\overline{x}=\mathrm{T}\mathrm{h}\mathrm{e}\mathrm{a}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}\mathrm{o}\mathrm{f}\mathrm{a}\mathrm{l}\mathrm{l}{x}_i$$

$$x_i=\mathrm{T}\mathrm{h}\mathrm{e}\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}\mathrm{o}\mathrm{f}\mathrm{e}\mathrm{a}\mathrm{c}\mathrm{h}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{p}\mathrm{o}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{t}$$

$$N=\mathrm{T}\mathrm{h}\mathrm{e}\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}\mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r}\mathrm{o}\mathrm{f}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{p}\mathrm{o}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{s}.$$

The Google Forms survey was distributed to 20 respondents, including both urban farmer testers and general users who had tested the HydroFarm application embedded within the form. This approach allowed respondents to evaluate the application from a UI/UX perspective. The accumulated responses were calculated using the average SUS formula, and the HydroFarm application received an average SUS score of 81.875. This score was then categorized based on its percentile range, as shown in Table 13.

Based on Table 13, HydroFarm received an A rating, as the average SUS score of 81.875 exceeds the threshold of 80.3. This score is further interpreted into three categories: acceptability, final grade scale, and adjective rating based on a scale of 0–100, with increments of 10. Figure 23 illustrates the correlation between the SUS score of 81.875 and these three categories.

In terms of acceptability range, the score of 81.875 falls into the "acceptable" category, as it exceeds the minimum acceptable average score. On the grade scale, the SUS score falls within the B grade range, as it lies between 80 and 90. Regarding the adjective rating, the score of 81.875 is classified as excellent, as it surpasses the threshold for excellence, which is set at 80. This indicates that HydroFarm is a well-designed, informative, and user-friendly application. Table 14 summarizes the evaluation results from the respondents.

Result of entire HydroFarm system

Based on a series of tests conducted on the HydroFarm system, it has successfully met its objectives and functions for monitoring and controlling hydroponic plants. In the IoT subsystem, the DHT22 and DS18B20 sensors do not require calibration since they are digital sensors, meaning they are already calibrated digitally. However, the TDS and pH sensors, being analog sensors, require calibration. When tested with calibration solutions, both sensors showed a high percentage of error, necessitating calibration using a linear regression approach. After calibration, the percentage of error for the TDS and pH sensors decreased to below 3%, indicating that the instruments are now providing accurate data and are ready for use. However, there is a delay of 5–7 s in the operation of the IoT devices, which is attributed to factors such as internet connection quality, server load, hardware performance, and communication methods. Efforts to address this issue have included replacing antennas with specifications of 3dBi, 6dBi, and 38dBi, as well as changing microcontrollers and internet connections, but the delay persists.

In the image processing subsystem, the model designed to detect the condition of hydroponic plants achieved an accuracy of 96%. Nonetheless, the model occasionally makes errors in detecting leaf conditions, leading to less accurate predictions. This issue is attributed to the limited dataset and variability in dataset acquisition techniques. Therefore, this model requires improvement to achieve better accuracy.

In the mobile application subsystem, there is a login feature using Google accounts, eliminating the need for manual registration. The HydroFarm application can monitor data from the DHT22, TDS, DS18B20, and pH sensors in real time using Firebase Realtime Database. The application also allows users to set nutritional targets on the automatic page, as well as manually control the pump and monitor its status through indicators. In the scanning feature, users can take pictures of leaves via the gallery or smartphone camera to predict whether the leaves are healthy or not using the CNN model. The profile feature includes information such as the user’s email username, a logout option, and a description of the HydroFarm application along with usage guidelines.

The evaluation results of the HydroFarm application using the System Usability Scale (SUS) from 20 respondents, consisting of expert testers and general users, show an average SUS score of 81.875. The application received an A rating because it exceeds 80.3, which falls into the “acceptable” category according to acceptability ranges. The application falls into grade B on the final score scale (80–90) and is classified as “excellent” in adjective rating since it exceeds the minimum threshold of 80. This indicates that the HydroFarm application is rated as a good, informative, and easy-to-use application by the respondents.