Search

Accurate and timely information about production estimates of wheat is useful for policymakers and government planners. The traditional methods for yield forecasting are labor insensitive and time-consuming therefore remote sensing is an effective approach for precise yield forecasting. The study was planned to develop a comprehensive framework for yield forecasting and to assess interannual yield variability in semi-arid regions. For wheat area classification, the peak season Landsat-8 satellite images were acquired, and Top of Atmospheric (TOA) correction was performed. The ground-truthing points of 100 farms were collected from the study area for the training of algorithms. The eight machine learning algorithms were used tune and tested using 10-k fold cross-validation and the best model was used for land cover classification of wheat. For yield forecasting, the temporal normalized difference vegetation index (NDVI) and land surface temperature (LST) were derived for the wheatgrowing season from November to April. A Principal Component Analysis (PCA) was used to variable selection and then Least Absolute Shrinkage Selection Operator (LASSO) analysis was performed to develop coefficients of the yield forecasting model. The developed model was further used in yield forecasting of 10 years (2008-2018) in four semi-arid regions. The predicted yield was compared with Crop Reporting Service (CRS), Pakistan department. The results of all machine learning algorithms showed an accuracy of 88% to 96%, however, the Random forest algorithm showed higher accuracy, which was further used for classification. The wheat estimated area of 6.9% was less than reported by CRS. For interannual variability, the relationship of observed (CRS) and predicted yield of 10 years showed a close relation with R2 ranged from 0.69 to 0.75 in the semi-arid region of Punjab, Pakistan. It was concluded that machine learning algorithms can be used as novel tools for yield forecasting and assessment of interannual yield variability. [ABSTRACT FROM AUTHOR]

Multi-Model Ensembles (MMEs) are often employed to reduce the uncertainties related to GCM simulations/projections. The objective of this study was to evaluate the performance of MMEs developed using machine learning (ML) algorithms with different combinations of GCMs ranked based on their performance and determine the optimum number of GCMs to be included in an MME. In this study ML algorithms; Artificial Neural Network (ANN), K-Nearest Neighbour (KNN), Support Vector Machine (SVM) and Relevance Vector Machine (RVM) were used to develop MMEs for annual, monsoon and winter; precipitation (P), maximum (T max) and minimum (T min) temperature over Pakistan using 36 Coupled Model Intercomparison Project Phase 5 GCMs. GCMs were ranked using Taylor Skill Score for individual seasons and variables, and then using a comprehensive Rating Metric (RM) overall rank of each GCM was determined. It was found that, HadGEM2-AO is the most skilled GCM and IPSL-CM5B-LR is the least skilled GCMs in simulating the 3 climate variables. The performance of MMEs did not improve after the inclusion of about 18 top-ranked GCMs. Thus, it was understood that the optimum performance of MMEs is achieved when about 50% of the top-ranked GCMs are used. The inter-comparison of MMEs developed with ANN, KNN, SVM and RVM revealed that KNN and RVM-based MMEs show better skills. It was found that RVM yields MMEs which show smaller variations in performance over space unlike ANN which displayed large fluctuations in performance over space. KNN and RVM are recommended over SVM and ANN for the development of MMEs over Pakistan. • Optimum performance of multi-model ensemble is achieved with 50% of top-ranked GCMs. • K -Nearest Neighbour and Relevance Vector Machine are good for multi-model ensembles. • Artificial Neural Network multi-model ensembles showed large performance fluctuations in space. • Machine learning-based multi-model ensembles outperformed simple ensmeble mean. [ABSTRACT FROM AUTHOR]