A Data Mining Driven Framework for Image-Based Crop Stress Analysis Using IPCA, FP-Growth, ConvLSTM, and Sequential Models

Abstract

Detecting crop stress before it becomes serious has a major impact on the food supply. It helps in avoiding the loss of crops, increases productivity and helps farmers to practice precision farming. Most technologies today are designed specifically to either detect crop stress or to predict how well the crop will grow. Many of the technologies do not combine the crop's spatial, spectral, and temporal data into a single framework by using an integrated assessment of the overall health of the crop. This paper presents a new framework, a data mining driven framework for image-based crop stress detection and yield prediction. The framework consists of three phases: 1) A new image processing framework utilizing Adaptive Wiener Filtering was developed for improving the quality of images by removing noise and maintaining the important features of the crops. 2): The Independent Principal Component Analysis (IPCA) method for extracting discriminative spectral and textural features from the images of the crops. 3) The association rule mining framework, FP-Growth, were utilized to determine the co-occurrence relationship of the extracted features. A Convolutional Long Short-Term Memory (ConvLSTM) network is then used to identify spatiotemporal patterns in order to classify the crop's stress anomalies. A Hidden Markov Model (HMM) was employed to measure the rate of progress of the crop stress over time, and Multi-Criteria Decision Analysis (MCDA) were used to calculate an overall score of the severity of the crop's stress level, thus enabling the farmer to make informed decisions. The development of this framework was based on an augmented dataset of wheat disease including 5000 images of the five specific classes of crop stress. Through the use of the framework, the results produced a classification accuracy of 98.9%, with an F1-score of 98.7%. The proposed framework outperformed other conventional classification models including CNN, Random Forest, LSTM, and SVM based on performance metrics.