A study on leaf area index and SAR image of oil palm with entropy decomposition and deep learning classification

In recent years, the increase in demand for food resources leads to the rapid expansion of oil palm plantations which is one of the more prominent commercial crops for countries located near the Equator. To cope with the expansion, a better plantation monitoring system is required which can be done via satellite remote sensing. Leaf area index (LAI) is one of the prominent biophysical properties of oil palms used by many to determine the overall health of oil palms and LAI is also monitored and maintained at a certain level to optimize fruit productivity. Therefore, it is important for plantation manager to monitor the LAI of plantation plots over time to perform optimization measures where microwave remote sensing can be used to aid them in monitoring their plantations. This paper attempts to explore the relationship between LAI of oil palm with microwave backscatter and later uses Deep Learning (DL) to classify the oil palm plots based on their LAI. Ground truth measurements were conducted in Lekir Estate, Perak state, Malaysia in October 2013 to obtain the vegetative parameters of oil palms together with C band microwave image from Radarsat-2 taken during the same time period. First, the relationship between LAI of oil palms and satellite data processed through entropy decomposition was studied in order to obtain a clearer picture on how LAI of oil palms affects microwave backscattering mechanism. Unsupervised Entropy Decomposition with Wishart Classifier was then used to cluster the data in order to observe the separability of the data in terms of LAI. Due to the constraint of C band data with fast attenuation of signal within the palm canopy, only 3 clusters can be obtained from the Lekir Estate microwave image which are ground, young palms (lesser than 3 years old) and mature palms (older than 3 years). Based on this information, ground truth information is then prepared into the three particular classes as training data for supervised classification. Deep Neural Network (DNN) via Caffe framework was used to perform the classification. In the first experiment, classification was done with entropy decomposition data where H and Alpha were used as input data for the DNN. DNN managed to obtain testing accuracy of 72.1%. To improve accuracy, another experiment using calibrated intensity data was used instead. DNN managed to obtain 83.33% testing accuracy for this experiment. This shows that modern machine learning classifier works better with raw data instead of processed data such as entropy decomposition as processed data have loss of potential information. Other than that, this work also found that better data separability in the context of LAI can be achieved if microwave with higher penetration power such as L band or P band to be used.