SmartHaven: AI-Enabled IoT System for Real-Time Monitoring of Poultry Health Through Vocalization Analysis and Environmental Sensing.

SmartHaven: AI-Enabled IoT System for Real-Time Monitoring of Poultry Health Through Vocalization Analysis and Environmental Sensing. Poultry farming plays a crucial role in food security and economic development across Sub-Saharan Africa, yet high mortality rates due to delayed disease detection, limited access to veterinary services, and inadequate infrastructure, which includes poor housing conditions for the birds, are some of the recurrent challenges that have discouraged a lot of local farmers from engaging in poultry farming. For instance, in Cameroon alone, about 98,000 tons of chickens, equivalent to 24 million birds, die yearly. Existing interventions, such as the SaveTheChicken mobile app, which aids in early disease detection through image analysis, have a few shortcomings. The image analysis takes about 15 minutes, and the predicted outcome has to be confirmed by a veterinarian before the final results are sent to the farmer. Though somewhat effective, the solution is time-consuming, and the prolonged decision-making process can delay timely interventions. This identified gap could be resolved through SmartHaven, an AI-enabled IoT system that combines real-time monitoring of poultry health through vocalization analysis with continuous environmental sensing on a single platform. An INMP441 microphone was used to record the audio of the chickens, while the DHT22 sensor recorded the temperature and humidity levels in the poultry farm. The audio data was saved on an SD card and, together with the environmental data, was sent through the ESP32 microcontroller to the Supabase database for storage. The captured audio was pre-processed by adding Gaussian noise, then converted into mel-spectrograms and fed into a fine-tuned, lightweight VGG-16 model trained on a poultry vocalization signal dataset, obtained from research conducted by a group of researchers at Bowen University. The training achieved an accuracy of 91.32% with a loss of 0.3489 and classified the results into two categories, either healthy or unhealthy. The model was deployed on the HuggingFace platform, and the API service was implemented using FastAPI to enable communication between the mobile app and the deployed model. The recorded audio was stored in Supabase and then sent through FastAPI for classification. The returned results, along with the sensor data, were displayed directly on the mobile app dashboard. The system was deployed on a poultry farm with 2,500 chicks in Nyamata, Bugesera, Rwanda, where it correctly distinguished between healthy and unhealthy vocalizations and flagged unfavourable conditions, that is, high or low temperature/humidity levels and an unhealthy poultry state, through automated alerts, thus eliminating farmers' prior practice of manual consultation and hourly temperature checks. The farmers found the solution to be innovative, with requests for the system to include gas/smoke sensing for air quality monitoring. This work demonstrates a low-cost, effective, deployable pathway using machine learning and IoT sensing for smallholder poultry farming in low-resource settings and also points to future integration in multiple farm settings and individual sick-bird identification using computer vision. Keywords: poultry health monitoring, acoustic classification, IoT sensing, mel-spectrograms, AI for agriculture, low-resource deployment.