Although land degradation (LD) is known as a severe environmental problem, spatial predictive modelling of this phenomenon remains a challenge. This research aimed to develop a new conceptual framework to predict LD susceptibility based on net primary production (NPP) and machine learning approaches. The annual NPP over the period 2001–2020 were obtained using MOD17A3 and the trend of NPP changes was considered to investigate the occurrence sites of LD within Qazvin Plain, in Qazvin Province, Iran, under a semiarid climate, with an area of about 9500 km2. An inventory map of LD was generated based on the LD study sites. The locations were randomly split‐sampled as training (70%) and testing (30%) datasets to evaluate the efficiency of the built models. Fifteen geo‐environmental factors were considered as LD predictive variables such as altitude, slope, land use, and temperature. Four advanced machine‐learning techniques were performed to model LD susceptibility. Finally, the predictive efficiency of the models was measured utilizing the area under the (ROC) curve Area Under the ROC Curve(AUC) and true skill as statics (TSS). The results indicated that the randomForest (RF), with the AUC = 0.81 and TSS = 0.5, showed the highest efficiency for predicting LD in the Qazvin Plain followed by boosted regression tree (BRT) with AUC = 0.76 and TSS = 0.47, support vector machine (SVM) with AUC = 0.71 and TSS = 0.39, and classification and regression tree (CART) with AUC = 0.63 and TSS = 0.31. The findings illustrated that altitude was the most influential variable within RF, BRT, and SVM while rainfall showed the most important contribution in modelling based on the CART algorithm. This study proposed a new modelling framework that is easily replicable in different contexts for the assessment of LD modelling and analysis.