Agroforestry Mapping using Multi Temporal Hybrid CNN+LSTM Framework with Landsat 8 Satellite Imagery and Google Earth Engine

Abstract

Agroforestry is indeed a traditional practice followed in tropical countries like India. About 28.43 million hectare area is used for agroforest cultivation. By 2050 India has the mission of increasing the area under agroforestry to 53 million hectares. In this study, we have made an effort to map the agroforest areas using the geospatial tools and hybrid deep learning techniques. The land utilized for cultivation and various agroforestry activities such as rubber, tea, coconut, and banana plantation were classified as forest canopy by the existing classifiers taking the tree canopy density as a parameter. In light of proposing a solution to the issue, we have put forth a multi temporal hybrid deep learning framework which is a fusion of convolutional neural network, a deep neural net and long short term memory network to classify agroforestry distinguishing it from the forest canopy using remote sensing data. The experimentation was carried out in the southern districts of India, and Landsat 8 imagery was used to classify the agroforestry of the study area that includes tea, banana, rubber, coconut, and crop lands. An efficient multi temporal hybrid deep learning framework was designed to classify the agroforest plantation distinguishing it from crop lands and forest clusters. The experimental results of multi temporal hybrid CNN+LSTM outperformed CNN, LSTM, BiLSTM model reducing the error rate with respective accuracy and kappa score of 98.23% and 0.88. The proposed method provides a benchmark to accurately classify and estimate the LULC, particularly mapping the agroforest plantation for other regions across the country.