Early detection of basal stem rot (BSR) disease in oil palm trees is important for the sustainable production of palm oil in the limited land for plantation in Southeast Asia. However, previous studies based on satellite and aircraft hyperspectral remote sensing could not discriminate oil palm trees in the early-stage of the BSR disease from healthy or late-stage trees. In this study, hyperspectral imaging of oil palm trees from an unmanned aerial vehicle (UAV) and machine learning using a random forest algorithm were employed for the classification of four infection categories of the BSR disease: healthy, early-stage, late-stage, and dead trees. A concentric disk segmentation was applied to tree crown segmentation at the sub-plant scale, and recursive feature elimination was used for feature selection. The results revealed that the classification performance for the early-stage trees is maximum at the specific tree crown segments, and only a few spectral bands in the red-edge region are sufficient to classify the infection categories. These findings will be useful for future UAV-based multispectral imaging to efficiently cover a wide area of oil palm plantations for the early detection of BSR disease.

Due to environmental threats and weather uncertainty concerns, oil palm yield prediction is crucial for sustaining crop production. This can be achieved through machine learning and utilising remotely sensed data to predict crop yield. However, the comparative studies on remotely sensed data in adopting the machine learning models are still limited due to the data accessibility. Therefore, we compare and evaluate the prediction accuracy between different satellites, namely MODIS and Landsat-7, using machine learning algorithms and the topology of deep neural networks. Random forest and stacking outperformed linear regression, ridge regression, and lasso regression for both Landsat-7 NDVI (R2= 0.78–0.80; RMSE=1.00- 1.26 tonnes per hectare; MAE=0.77- 0.79 tonnes per hectares; MAPE=0.03-0.04 tonnes per hectare) and MODIS NDVI (R2= 0.60–0.65 tonnes per hectares; RMSE= 2.72–2.81 tonne per hectares; MAE= 1.42-1.55, MAPE= 1.01- 1.02 tonnes per hectares). The Landsat-7 NDVI revealed that neural networks with a deeper network topology (R2= 0.85; RMSE= 1.42 tonnes per hectare; MAE=0.57 tonnes per hectares; MAPE=0.06 tonnes per hectare) outperformed neural networks with a baseline and broader network topologies in terms of performance. In contrast, MODIS-NDVI revealed that the neural network with a wider network topology had the highest overall prediction accuracy and the lowest prediction error (R2= 0.75; RMSE= 2.81 tonnes per hectare; MAE=2.27 tonnes per tonnes; MAPE= 0.13). Because of its higher spatial resolution in comparison to MODIS, landsat-7 NDVI used in neural networks with a deep network topology provided the best model performance. Although the use of NDVI as a single input factor may cause uncertainty in some extents, it is an efficient and reliable method for improving yield estimation with the use of medium-resolution satellites, which has important implications for early warning towards the reduction in yield production.

Recent advances in land use and land cover (LULC) classification with remote sensing imagery are driven by state-of-the-art models such as Convolutional Neural Networks (CNNs). Advanced CNN architecture like ResNet can enhance overall classification performance by incorporating residual skip connections. The integration of 3D feature extraction and ResNet architecture suggests a potential improvement in classification tasks. This paper explores the potential of the 3D ResNet model for LULC classification, comparing it with baseline approaches (Support Vector Machine, Random Forest, XGBoost, 1D CNN, 3D CNN) and state-of-the-art 3D models (3D VGG, 3D DenseNet) using WorldView-2 satellite imagery. The 3D ResNet-18 model, fine-tuned via transfer learning on multispectral images, demonstrates significant improvements in classification performance over machine learning models. It achieves the highest Overall Accuracy (OA) of 99.66% and Kappa Accuracy (KA) of 99.39% on the primary dataset. Despite having slightly lower performance on the external validation dataset (OA:82.89%, KA:80.05%) than 3D DenseNet, it is highly efficient with processing times of 490.2 minutes and 3.6 minutes for both datasets respectively. McNemar’s test results show 3D ResNet and 3D DenseNet have significant differences in classification performance (p<0.05) against other models consistently for both datasets.