Fusion of IoT Sensor Data and Image Processing for Comprehensive Crop Health Assessment

Abstract

Effective crop health monitoring requires the integration of heterogeneous data sources that capture both environmental conditions and crop-level responses. Conventional single-modality approaches, relying either on in-ground sensor measurements or aerial imagery in isolation, fail to exploit the complementary strengths of each technique, resulting in limited diagnostic accuracy and delayed intervention. This study proposes an integrated approach that fuses data acquired by IoT-connected sensor networks with multispectral images captured by unmanned aerial vehicles (UAVs), enabling comprehensive crop health assessment across an entire field. Consider a system that deploys a network of sensors throughout a 2-hectare area to continuously watch the moisture and temperature, as well as other critical parameters, such as electrical conductivity, atmospheric humidity as well as the nutrient level in the soil. Simultaneously, a special UAV, a drone known as the DJI Phantom 4 Multispectral, take pictures of the field at five different spectral frequencies. However, the most interesting thing is the following: to do all this, the system relies on a special type of artificial intelligence known as a hybrid deep learning architecture. It resembles a twostep procedure, where the one section analyses the trends in the sensor data across time and is called one dimensional convolutional neural network, whereas the other section uses a pre-trained version of ResNet-50 and is responsible of making significant features of the images captured by the drone. Then, it is all unified with an eight-head multi-head attention mechanism, taking all the various kinds of data and forming a bigger picture. This will enable the system to memorize and establish relationships among the various kinds of data, which forms a potent source of knowledge and management of the field. The synchronized dataset comprised 2,520 data points collected over 120 days (April–August 2024), with 103 days of active data capture. The proposed hybrid deep learning architecture achieved a crop health classification accuracy of 94.5%, compared to 80% for conventional single-modality methods — a statistically significant improvement of 14.5 percentage points. The system classifies crop status into five categories: healthy, water stress, nutrient deficiency, disease, and pest infestation. The results demonstrate that cross-modal IoT–image fusion delivers earlier, more reliable diagnosis of crop stress conditions, enabling data-driven farm management decisions that support precision agriculture at scale.

Plant phenotyping relevance

UAVマルチスペクトル画像とIoTセンサーデータを融合し、作物の健康状態・ストレス状態を分類する深層学習手法が研究の中心であり、植物状態の取得・推定方法として適格。

Code and data availability

The supplied blocks describe a field study (IoT sensors, UAV multispectral imagery, hybrid CNN fusion model) but contain no data availability statement, no public dataset deposit of the authors' sensor/image data, and no code or model availability language or URLs. The only public resource mentioned is the PlantVillage

No evidence-backed public reproduction asset is currently recorded.