Deep Learning Approach for Detection and Prediction of Pest Infections On Plants in Greenhouses

Abstract

Pest infestations remain a serious threat to greenhouse agriculture ultimately resulting in reduced yields, increased cost of production, and food safety concerns. Traditional pest monitoring strategies are predominantly manual in nature, which can be laborious, time-consuming, and exhibit human error. These downfalls can affect timely action on issues and a positive relationship management with pest monitoring can mean excessive pesticide use with financial and/or environmental impacts. This paper proposed an AI-enabled hybrid deep learning model to automate pest detection and estimate the probability of a pest outbreak. The hybrid model combined Convolutional Neural Networks (CNNs) for spatial (image-based) pest identification with Long Short-Term Memory (LSTM) networks to estimate probabilities of outbreaks based on sequences of environmental ((local) environmental data temperature, humidity, as well as, recorded counts of pests) and pest counts to help improve timing of interventions and proactive pest management solutions. The study made use of both primary and secondary datasets. The primary dataset was collected from three greenhouses in Limuru, Naivasha, and Thika, Kenya as well as high-resolution crop images (48 mega-pixels) representing various pest infections. Local environmental data for temperature, humidity, and pest counts were collected from the greenhouses to use as sequential variables in the LSTM component of the workflow and assist with estimating the accuracy of predicted outbreaks. Secondary datasets of pre-annotated pest images and historical climate records, including PlantVillage and IP102, provided bulked training data that improved models’ robustness and generalizability to unseen datasets. The researcher ensured stratified sampling to capture representation of all greenhouse types, farm sizes, and agro-climatic conditions. As part of preprocessing, all datasets underwent the following steps: image augmentation, noise removal and feature normalization. All model training, including hyperparameter tuning, occurred in a GPU-enabled Google Colab environment, and early stopping was used to avoid overfitting. The hybrid CNN-LSTM model produced a classification accuracy of 94.7%, precision of 93.9%, recall of 93.8%, and F1-score of 93.2%. The time-series predictive component of the LSTM model produced strong predictive performance with a mean absolute error (MAE) of 0.14, and R² value of 0.89. This showed that the environmental and measured sequence data of pests improve outbreak prediction. It was shown that the hybrid model accurately identifies pest infections, as well as predicting their outbreaks, which will contribute to providing early and timely interventions in the pest control system. Both high-resolution image data and measurements of local environments support a scalable and resilient choice for greenhouse pest management - the use of less pesticides, leading to sustainability in agriculture. In conclusion, this study has demonstrated the effectiveness of a hybrid deep learning framework for integrated pest management with operational and financial implications.