Detecting <scp> <i>Pteridium arachnoideum</i> </scp> and <scp> <i>Tithonia diversifolia</i> </scp> in a Multiclass Land‐Cover Framework in the Cerrado: A Deep Learning Time‐Series Benchmark With Spatial Cross‐Validation

ABSTRACT This study presents the first deep‐learning time‐series benchmark for detecting two encroaching species in the Brazilian Cerrado: the native super‐dominant fern Pteridium arachnoideum and the invasive alien shrub Tithonia diversifolia . We evaluate 43 architectures across six families (CNNs, RNNs, Transformers, and three hybrid groups) on monthly Planet NICFI satellite imagery in Brazil's Federal District, adopting a 13‐class land‐use/land‐cover scheme with 16,250 samples. CNNs and hybrid CNN + RNN models achieve the highest accuracy, with FCN and LSTM‐FCN emerging as the most robust architectures (F1 &gt; 98% under random splitting, ∼82% under spatial cross‐validation). Tithonia diversifolia is reliably detected across spatially independent regions owing to its distinctive flowering phenology (F1 &gt; 95%), whereas P. arachnoideum exhibits greater spatial variability (F1 &gt; 76%) due to spectral overlap with forest formations. Wall‐to‐wall classification maps confirm ecologically coherent spatial distributions for both species. Spatially independent five‐fold cross‐validation reveals that conventional random splitting inflates performance by 16–25 percentage points across all architecture groups. These results provide practical guidance for selecting architectures suited to operational monitoring and underscore the importance of spatially independent evaluation in remote‐sensing classification.