A novel pixel-based deep neural network in posterior probability space for the detection of agriculture changes using remote sensing data

Abstract

Agricultural land classification is a crucial and demanding task, essential for managing resources and tracking changes in farming activities. Remote sensing (RS) is an excellent technology for monitoring agricultural land and detecting seasonal fluctuations globally. Deep learning models offer promising prospects for crop monitoring. While traditional machine learning methods struggle to capture temporal variations in agricultural land efficiently. This study addresses the challenge of accurately classifying land cover and detecting year-to-year changes using a deep learning (DL)-based approach. The novelty of this research is an integration using a pixel-based deep neural network (PDNN) classifier, which will advance the classification abilities for identifying land cover classes. By comparing images taken over time, the PDNN can help identify different land cover efficiently. The thematic images using the PDNN were derived, and change detection was carried out by adopting a posterior probability space (PPS)-based change detection. The application of the proposed model is demonstrated using the Landsat-9 dataset over Moga District, Punjab, India. Compared to the random forest (RF) and support vector machine (SVM), the PDNN model achieved great performance. While PDNN had an accuracy ranging from 90.6 % to 93.6 %, RF and SVM had lower accuracies, with RF ranging between 86.8 % and 92.2 % and SVM between 88 % and 92.4 %. The PDNN model also excelled in detecting changes in land cover, showing an accuracy between 87.4 % and 90 %, while RF achieved 82.9 %–86.2 % and SVM ranged from 79 % to 83.9 %. The proposed model was adept at capturing changes in agricultural land cover, such as year-to-year variations. The PDNN model demonstrated superior proficiency in capturing seasonal and year-to-year variations in agricultural land cover, effectively identifying subtle transitions in crop cycles. This highlights its potential for long-term agricultural monitoring and precision farming applications. This approach would serve as a key to sustainable agriculture, which guides farmers and policymakers to make better choices.