Evaluating the spatial–temporal transferability of models for agricultural land cover mapping using Landsat archive

Changes in policy and new plans can significantly influence land use and trigger land use change in the long term. The data for pre- and post-policy implementation is necessary to assess the specific policy’s impact on land use. In the early nineties, Germany started promoting renewable energy production, including bioenergy, which changed the agricultural landscape. Remote sensing (RS) image-based machine learning models can be beneficial for mapping agricultural land use in the present and the past. However, machine learning classification models trained on RS data from specific training sites and time may not be able to predict data for unknown sites and unknown temporal points due to changes in crop phenology, field features, or ecological site circumstances because most of the models are limited in their performances according to variations of the training data set. Therefore, this study aims to assess the spatial–temporal transferability of Landsat-based agricultural land use type classification. The study was developed to map agricultural land cover (5 classes: maize, grasslands, summer crops, winter crops, and mixed crops) in two regions in Germany (North Hesse and Weser-Ems) between the years 2010 and 2018 using Landsat archive data (i.e., Landsat 5, 7, and 8). Two machine learning models (random forest − RF and 2D convolution neural network – 2DCNN) were trained and evaluated according to no transferability (reference) scenario and three spatial–temporal scenarios using mF1 and class level F1 values. Three model transferability scenarios were evaluated: a) temporal – S1, b) spatial – S2, and c) spatiotemporal – S3. The reference scenario, without transferability, achieved an overall accuracy of 89.1% and a macro F1 score of 0.74 for RF and 89.9% and 0.75 for CNN, respectively. Under three transferability scenarios (S1, S2, and S3), the macro F1 scores decreased to 0.67, 0.66, and 0.62 for RF, and 0.68, 0.62, and 0.58 for CNN, respectively. The dissimilarity between the data employed to train the model and data from the new domain indicated a clear link that could explain the reduction in model predictability. Moreover, the performance degradation could be attributed to the disparity in environmental, climatic, and crop calendar conditions between the two domains. Understanding the extent of model performance degradation during transferability is crucial for developing effective strategies to mitigate these issues and enhance the generalisability of machine learning models for agriculture land cover mapping.