Integrated design of an efficient multi spectral imaging and federated learning framework for precision crop disease diagnosis in low-resource farming communities

In this study, the experimental setup was designed to robustly test the multi-modal machine learning-supported crop disease diagnosis framework under conditions that represent low-resource farming environments. The major dataset used for the experiment is hyperspectral images of a variety of crops collected using portable hyperspectral cameras that scan the spectrum from 400 to 1000 nm at 10 nm resolution. These images were complemented by RGB images captured by commercial-grade cameras, the environmental conditions of temperature, humidity, and soil moisture measurements from IoT sensors, and the reports from farmers about disease symptoms and results of treatments. The dataset consists of approximately 10,000 hyperspectral image cubes, 20,000 RGB images, and their environmental measurements, all covering more than 15 crop varieties and 20 disease categories. Data were collected over a period of 6 months over the three geographically diverse farming regions to ensure variability in crop species, environments, and disease prevalence. Each hyperspectral image cube had dimensions (H × W × B) (256 × 256 × 60), where ‘B’ corresponds to spectral bands, providing fine-grained spectral information for each set of spatial pixels. To mimic resource-scarce environments, light denoising algorithms were applied along with spectral bands selection during data preprocessing, which was done to reduce input space while preserving the disease-specific critical features. The 3D-SSCNN extracted spatial and spectral features from the preprocessed hyperspectral data, while the MTAN fused this data with RGB images and environmental metrics to generate these integrated stress profiles. Model training was done using federated learning, where the systems by farmers were used as decentralized nodes. Each node trained locally for five epochs at a learning rate of 0.01, and model gradients were uploaded securely to a central server for global aggregation. To test the TP-LSTM, sequential data in the form of 10-day disease progression observations were used, where one day was taken as the temporal resolution. The initial treatment policies for reinforcement learning in RL-FO were derived from expert-curated guidelines, with rewards calculated based on yield improvement of up to 15% and cost minimization targeting 20–30%. The study utilized the Plant Village Dataset and supplemental environmental data retrieved through public repositories, such as NASA EarthData and SoilGrids. The Plant Village Dataset consists of more than 50,000 labeled RGB images covering 38 crop species and 26 distinct disease classes, such as powdery mildew, late blight, and leaf rust. Such high-quality images are augmented to represent various environmental conditions, including variability in lighting and partial occlusions. Hyperspectral data were synthesized using the spectral response curves provided by agricultural research centers, and spectral samples from fields were added to the dataset through crop disease hyperspectral datasets published by agro-institutions, thus including nearly the entire range of spectral bands-the 400–1000 nm range-critical for disease biochemical marker detection. Secondly, IoT environmental data from NASA EarthData and SoilGrids at high resolutions of measurements for soil attributes (e.g., pH, nitrogen level) and atmospheric variables (e.g., temperature, humidity, and precipitation) has been collected. The datasets are then combined and spatially and temporally aligned to form a multimodal dataset of diverse agricultural conditions. The use of real-world and augmented data ensured that the proposed framework was robust, thereby including variability in disease appearance, crop types, and environmental stresses critical for model generalizations (Fig. 3).

Since the contextual dataset exhibited tremendous variability with diseases like wheat leaf rust, rice blast and tomato late blight being well represented and abiotic stresses, including drought and salinity, evaluation metrics involved classification accuracy, training latency, prediction latency, and even a model’s ability to be robust under noisy data conditions. For instance, 3D-SSCNN resulted in 95% accuracy with a processing time of 0.3 s per hyperspectral cube. Fed-DiagNet decreased training time by 40% compared to the centralized models while achieving 92% accuracy across the regions. In precision, TP-LSTM was able to predict the amount of disease severity up to a 10-day horizon with an 85% precision. The treatment recommendations by RL-FO were 88%, and within the simulated trials, 80% of the adoption occurred among farmers. This experimental setup has therefore underscored the scalability, precision, and adaptability of the framework, demonstrating its potential for the transformation of disease management in low-resource agricultural contexts. The experimental results emphasize the framework proposed in generalizing better overall compared to other aspects such as classification accuracy, disease latency detection, accuracy in temporal prediction, performance in federated learning, efficiency in recommending treatments, and the process of identifying stress. This paper mainly includes comparisons of the proposed approach with both conventional and recent state-of-the-art methods, including Method⁵, Method⁸, Method¹⁸, the Swin Transformer with federated learning approach²⁰, and the Lightweight DenseNet framework²⁵, thereby enabling a comprehensive evaluation across hybrid, transformer-based, and computationally efficient diagnostic paradigms (Table 2).

All datasets showed significantly high classification accuracy, with a maximum of 96.3% in the multimodal setup. By integrating hyperspectral data, RGB images, and environmental information with the help of MTAN, the model catches disease-specific features better than others. The outcomes of Method⁸ and Method⁵ are lagging because of their inability to process spectral-spatial data and integrate multimodal information sets. The actual impact of this result can be critical for applications since higher classification accuracy can directly translate to fewer false diagnoses and drastically reduce unnecessary treatments and associated costs (Fig. 4, Table 3).

The proposed model also indicates that the data is processed much faster than the comparing methods since its latency was measured at 0.35 s under the multimodal setting. Method¹⁸ had the highest latency because it could not handle complex data efficiently due to the simplified feature extraction process that was applied. The low latency of the proposed model ensures real-time diagnosis, which is one thing so basically required for in-field applications where farmers need immediate feedback to make timely decisions (Fig. 5, Table 4).

The TP-LSTM, by showing consistent time horizon performance with 96.5% of accuracy in the one-day horizon, was able to keep its accuracy to a level of 85.4% even up to the 10th day mark. However, Methods⁵ and ¹⁸ have demonstrated extreme reductions in accuracies for extended horizons due to the lack of any advanced temporal modeling process. Superior temporal precision with the support of TP-LSTM eventually impacts the management of disease through preventive interventions as per the forecasted severity levels, which enables farmers to handle possible losses effectively for the process (Table 5).

In the proposed framework, Fed-DiagNet achieved a regional accuracy of 92.0% while reducing the training time by 40.2% compared to conventional centralized training approaches, where all multimodal data are aggregated and processed at a central server under identical training configurations. This improvement demonstrates the scalability and computational efficiency of the proposed federated framework. In contrast, Method⁵ and Method¹⁸ relied on centralized training, resulting in higher communication overhead, increased latency, and reduced adaptability across geographically distributed agricultural regions. The results presented above further prove the applicability of the proposed framework under geographically dispersed agricultural settings with high accuracy and with good computational efficiency levels (Fig. 6, Table 6).

Optimization of treatment strategies through the RL-FO system with 88.2% accuracy attained adoption among 81.4% of farmers. Lower adoption rates are, in fact reflected as that of Method⁵ and Method¹⁸, making use of simplified recommendation systems since their recommendations are less relevant and not cost effective for different scenarios. The higher adoption rate also reflects practical applicability by the RL-FO component because the implementation possibilities for such recommendations is more specific to the needs of a farmer, thus promoting confidence among farmers in their implementation process (Table 7).

The MTAN module was very accurate in detecting combined biotic and abiotic stresses at 92.6% whereas it outperformed all comparison methods. Method⁸ only performs moderately and does not reach the level of integration obtained in the MTAN module, and Method¹⁸ was highly challenged by the combination of stresses. These early intervention strategies then depend on accurate stress detection to address the right measures, which can enhance crop resilience and raise the levels of yield. Overall, these results demonstrate the great potential of this proposed model in revolutionizing precision agriculture cases. Its high precision, efficiency, and adaptability towards the resolution of considerable issues in crop disease diagnosis and treatment in the low-resource farming environment make for its strength levels. We now give an iterative validation use case for the proposed model process. This should enable readers to delve deeper into the full process.

The proposed framework against recent state-of-the-art developments is shown in Table 8. Transformer-based approaches such as Swin Transformer with federated learning²⁰ demonstrated improved feature representation and privacy preservation, while lightweight architectures such as DenseNet²⁵ achieved lower computational complexity. Semi-supervised ensemble frameworks³⁹ improved robustness under limited labeled datasets. However, these approaches do not simultaneously address multimodal data fusion, temporal disease progression modeling, privacy-preserving federated optimization, and adaptive treatment recommendation. As observed in Table 8, the proposed framework achieves superior diagnostic performance while maintaining lower inference latency and broader functional capability for real-time precision agriculture applications.

Beyond quantitative improvements in classification accuracy, latency, and treatment optimization, the proposed framework also provides important qualitative insights into crop disease progression and stress characterization. The spectral attention mechanism embedded within the 3D-SSCNN assigns higher weights to disease-sensitive wavelength bands, enabling identification of critical biochemical markers associated with crop stress conditions. Similarly, the cross-modal attention mechanism in MTAN facilitates adaptive fusion of hyperspectral, RGB, and environmental features, allowing the framework to capture both biotic and abiotic stress interactions under heterogeneous field conditions. The TP-LSTM module further provides interpretable temporal progression patterns by forecasting disease severity across a 10-day prediction horizon, thereby supporting proactive intervention strategies. In addition, the RL-FO module continuously refines treatment recommendations based on reward-driven feedback, enabling adaptive decision-making under dynamic agricultural environments. These characteristics collectively provide implicit interpretability and practical decision support for real-world precision agriculture applications.

Validation using iterative practical use case scenario analysis

In this section, outputs from all modules of the proposed architecture are assessed for their contributions by the MTAN, 3D-SSCNN, Fed-DiagNet, TP-LSTM, and RL-FO. The results highlighted the following: precise, actionable insights generated in process. For the case study, actual usage data was taken from the PlantVillage Dataset; more than 50,000 labeled RGB images of crops are available including for wheat and their respective diseases such as leaf rust, powdery mildew, and blight. Spectral data of wheat diseases response was scaled from the CropDisease Hyperspectral Dataset providing spectral cubes across 60 bands (400–1000 nm). Environmental data, such as soil pH, temperature, and humidity, were used from NASA EarthData and SoilGrids for high-resolution spatial and temporal measurements. For the analysis, each sample was a composite of RGB images, like the high-resolution images of rust-affected leaves, hyperspectral data cubes that captured the characteristics of leaf reflectance, and environmental data samples, which were recorded daily for 10 consecutive days. This multimodal dataset enabled the framework to quantify, identify, and predict disease progression and stress indicators in a very comprehensive manner, mimicking real-world field conditions for the diagnostic and management process of wheat leaf rust. The MTAN combines spectral, RGB, and environmental data into one representation of integrated stress. Table 9 summarises the features extracted from the multimodal dataset, along with their respective probabilities of stress.

The MTAN integrates multimodal inputs precisely, emphasizing biotic stress signals relevant to the leaf rust that has been detected. A combined stress probability of 94.7% success rate indicates its feasibility in the diagnosis of disease. The 3D-SSCNN extracts spectral-spatial features from hyperspectral data and infers disease-specific patterns. Table 10. A feature contribution with its corresponding classification outcomes.

The spectral-spatial attention mechanism gives higher weights to the critical bands, and hence, the contribution of the classification will be 95.0% for the leaf rust disease. The Fed-DiagNet aggregates the local model updates from three regional nodes. Table 11 exhibits the local accuracies and the global performance after the aggregation process.

The federated approach achieved global accuracy at a level of 92.0% with good balance between the combination and preservation of regional data privacy levels. The TP-LSTM makes predictions about the severity level of disease over a horizon of 10 days during the operations. Table 12 lists out the predicted levels on each day in the aggregated process (Fig. 7).

The model captures the severity of leaf rust over time accurately, so there is timeliness in intervention operations. The RL-FO system gives optimized treatment recommendations based on farmer feedbacks. Table 13: selected actions and their corresponding output (Fig. 8).

The combined treatment gave the highest reward and adoption rate, optimizing yield improvement and cost-effectiveness. Table 14: summary of the final diagnostic and decision-support outputs for the use cases.

The final results demonstrate the feasibility of the developed framework toward achieving a goal in real-time efficiency, besides high adoption by farmers through accurate diagnosis and actionable recommendations. These results underpin the possibility of the framework to transform the management of disease in precision agriculture scenarios.