Machine learning classification of macronutrient levels in oil palm trees with Landsat-8 imagery

Optimized fertilizer application is essential in ensuring the best yields and profits. In the last two decades, fertilizer application in the palm oil industry of Malaysia has increased by 169% and more than 200% in amount and value respectively. The excessive use of fertilizer is seen as a deterrent to sustainable agriculture, both environmentally and economically. Unfortunately, current nutrient monitoring methods for better fertilizer management are inefficient, desctructive and only adoptable for industrial plantations. This suggests the need for a more conventional and efficient approach for fertilizer management. Remote sensing (RS) and machine learning (ML) have provided possibilities to monitor crop nutrient in a more efficient approach. Therefore, this study aims to assess the feasibility of Landsat-8 OLI/TIRS RS imageries in classifying oil palm tree macronutrient levels on a plot basis with ML models such as Support Vector Machine (SVM), Multilayer Perceptron (MLP) and Random Forest (RF). The study consisted of 36 plots in which fertilizer applications and foliar analyses of palm leaves from Frond 17 (i.e. Nitrogen(N), Phosphorus(P), Potassium(K), Magnesium(Mg) and Calcium(Ca) ) were conducted for 5 consecutive years (2013 - 2017). Plot surface reflectance was acquired from specific scenes of respective years by applying filters, overlay extraction and vegetation indices (VI). The study explored the effects of the applied methodology with four scenarios as input for the models via a 50:50 calibration and validation data split for 30 iterations: 1.) initial spectral data; 2.) filtered spectral data; 3.) filtered spectral features with selection; and 4.) all filtered spectral features. Using the Jeffries- Matusita (J-M) distance, Rank filter was found to be the best filter for image filtering. VIs related to soil background or atmospheric correction were selected as predictors for N (i.e. EVI, SARVI, ARVI, GARI, MSAVI, EVI2), while those related to NIR for K (i.e. NDVI, TVI, IPVI). The best mean classification accuracy for N, K, Mg and Ca of each model at validation stage are as follow: SVM at 79.7%, 76.6%, 63.5% and 87% respectively; MLP at 76.1%, 73.4%, 61% and 87% respectively; and RF at 76.9%, 73.6%, 62.1% and 86.4% respectively. SVM, RF and MLP experienced improvement in performance with the use of filters or feature selection, with MLP benefiting the most. with the use of filters or feature selection or both (Scenario 2 and 3). However, uneven sample distribution led to misleading results for Ca and occasional overfitting by models for other nutrients. SVM was reported with the best model performance, followed by RF and MLP; while RF for model stability, followed by SVM and MLP. Accounting both factors, the study concluded the potential of applying Landsat-8 images in classifying plot N level excessiveness with RF while possibly K level excessiveness with SVM or RF.

Text FK 2021 60 IR.pdf Download (1MB)

Actions (login required)

View Item