Temporal scalability analysis of spectral remote sensing and Machine Learning-Based models for forest mapping in the tropical region of Nigeria

• Machine learning models trained and tested on January 2022 data reached a 0.93 F1-score for forest classification. • Models trained and tested on December 2022 data achieved a 0.91 F1-score for forest classification. • Multi-temporal models trained on January and December 2022 data showed reduced accuracy, with F1-scores of 0.74 and 0.71. • Models trained on Jan 2022 data showed poor performance on Dec 2022 data, indicating weak temporal scalability. The integration of machine learning (ML) algorithms with remote sensing (RS) data offers significant potential for enhancing dynamic forest characterization and change detection through high temporal density mapping. This study evaluates the effectiveness of RS-ML models for forest mapping in parts of southern Nigeria, with particular emphasis on assessing their temporal scalability using multi-sensor data from Sentinel-2 and Landsat 8 platforms. Three strategies were evaluated: the first involved training RS-ML model on spectral vegetation indices generated based on datasets from single date and applying it to datasets captured in different dates; the second entailed training distinct models for each datasets captured in certain time period and the third one utilized datasets from various dates (multi-temporal) to land cover mapping. Accuracy assessments of the land cover maps generated by all strategies were conducted to determine the loss of accuracy when a RS-ML model is trained and applied to another temporal dataset. The highest F1-scores for the forest class were obtained from ML-RS models trained and applied on the datasets captured in the same time period. For instance, when trained independently on datasets from January and December 2022, the model achieved F1-scores of 93.09% and 91.56% for the forest class, respectively. Conversely, a ML-RS- model trained based on both multi-temporal datasets − January and December 2022 jointly − registered a lower F1-score of 74.29% for the forest class. A similar decrease in accuracy was observed with a another multi-temporal (multi-season) model trained jointly on data from December 2022 and December 2023, resulting in an F1-score of 71.58%. This indicates a significant drop in accuracy when multi-temporal features are used for forest mapping. Thus, for further analysis of forest change mapping, the most accurate forest maps were utilized generated by using models with the highest F1-scores. Although the ML-RS models are highly accurate for specific dates, their temporal scalability poses a challenge propably due to the specific tropical conditions. Therefore, for land cover mapping it is crucial to select models that align with the specific temporal and environmental conditions and provide high accuracy. To address this challenge, the study recommends prioritizing adaptive modeling strategies. These strategies involve continuously updating and retraining ML-RS models to ensure they remain accurate over different periods, thereby improving the reliability of forest mapping over time. This approach emphasizes the importance of model selection based on the intended application period and environmental variability, to ensure the generation of accurate and reliable forest maps.