Hybrid deep neural network and random forest ensemble for multi-crop disease diagnosis and agro-risk prediction using weather, soil, and imaging datasets

Abstract

The early identification of crop diseases and correct agro-risk analysis are critical to sustainable agricultural productiveness and worldwide food security particularly at a time when climate change and environmental stressors are rising. In the following paper, the authors present the mentioned sophisticated hybrid framework, HDNN-RF FusionNet, which combines the best of both deep and ensemble learning when it comes to solving the task of multi-crop diseases diagnosis and agro-risk prediction on multimodal data. The framework can handle and train on a wide range of agricultural inputs such as satellite-based spectral images, drone-based spectral images, weather time series and profiles of soil composition. HDNN-RF FusionNet uses the multi-branch deep neural network architecture organization each of which branches is specialized to encode a particular modality. The imaging division receives the disease symptoms in the leaf textures and canopy structure in the spectral imagery using a convolutional neural network (CNN). The weather branch utilizes long short-term memory (LSTM) networks to identify time series aspects of changes in climate conditions affecting disease developments. The soil branch transposes a multilayer perceptron (MLP) to classify unchanging geographical characteristics, including nutrient concentrates, pH and humidity. All three branches output are fused through an attention-weighted fusion layer in which the most significant modality becomes prioritized dynamically according to the crop type and environmental condition. The fused feature vector is provided as an input to a Random Forest (RF) model, which has two tasks: (1) multilabel classification of crop diseases and (2) regression-based prediction of an agro-risk index representing the possible impact on yields under the existing and the predicted conditions. This mixed architecture is robust, interpretable, and well-generalized in regions and seasonality. The effectiveness of the model in early disease detection and risk prediction has been proved by running experimental analyses in real-world data, multi-seasonal and multi-location data. HDNN-RF FusionNet proposed is a scalable, smart, and smarting answer to next-generation agro-intelligent systems. The suggested HDNN-RF FusionNet had an overall accuracy of 96.4%, surpassing all benchmark models in multi-crop disease detection and agro-risk prediction.